Skip to content

Add validity checks for MoE FlatMM scatter and enable bf16 hardware atomic-add #3236

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
wants to merge 5 commits into
base: develop
Choose a base branch
from
+10 −6
Open

Add validity checks for MoE FlatMM scatter and enable bf16 hardware atomic-add #3236

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
5bfda1f
Add validity checks for MoE FlatMM scatter and enable bf16 hardware a…
msaffari-amd Nov 19, 2025
5b4ee55
correct clang-format
msaffari-amd Nov 19, 2025
8d56b0a
removed unused rtol_atol variable from example code
msaffari-amd Nov 21, 2025
122a981
clang format correction
msaffari-amd Nov 21, 2025
caf82d1
remove unused varable max_accumulated_value from example
msaffari-amd Nov 21, 2025
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
4 changes: 0 additions & 4 deletions example/ck_tile/18_flatmm/run_moe_flatmm_example.inc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -302,10 +302,6 @@ int run_moe_gemm_example_with_layouts(int argc,
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer()),
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer()));

const float max_accumulated_value =
*std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
K, 1 /*kbatch*/, max_accumulated_value);
c_m_n_ref_buf->FromDevice(c_m_n_host_ref.data());

const float rtol = std::is_same_v<ADataType, ck_tile::half_t> && IsInputGemm ? 1e-3 : 1e-2;
Expand Down
4 changes: 4 additions & 0 deletions include/ck_tile/core/arch/generic_memory_space_atomic.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,9 @@ CK_TILE_DEVICE void atomic_add(X* p_dst, const X& x);
template <>
CK_TILE_DEVICE void atomic_add<bf16x2_t>(bf16x2_t* p_dst, const bf16x2_t& x)
{
#if HAS_GLOBAL_ATOMIC_PK_ADD_BUILTIN
__builtin_amdgcn_global_atomic_fadd_v2bf16(c_style_pointer_cast<bf16x2_t*>(p_dst), x);
#else
union U32BF162_ADDR
{
uint32_t* u32_a;
Expand All @@ -128,6 +131,7 @@ CK_TILE_DEVICE void atomic_add<bf16x2_t>(bf16x2_t* p_dst, const bf16x2_t& x)
new_v = new_.u32;
cur_v.u32 = atomicCAS(dword_addr.u32_a, old_v, new_v);
} while(cur_v.u32 != old_v);
#endif
}

template <>
Expand Down
8 changes: 6 additions & 2 deletions include/ck_tile/ops/flatmm/kernel/moe_flatmm_kernel.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -623,7 +623,7 @@ struct MoeFlatmmKernel
{
return make_naive_tensor_view<address_space_enum::global, DstInMemOp>(
e_ptr,
make_tuple(IsInputGemm ? kargs.NumTokens * kargs.TopK : kargs.NumToken,
make_tuple(IsInputGemm ? kargs.NumTokens * kargs.TopK : kargs.NumTokens,
IsGateUp ? kargs.N / 2 : kargs.N),
make_tuple(1, kargs.stride_C),
number<1>{},
Expand Down Expand Up @@ -1250,6 +1250,8 @@ struct MoeFlatmmKernel
constexpr int MPerThread = TileEncodingPattern::Y2;
statically_indexed_array<statically_indexed_array<index_t, MPerThread>, NumMEpiTile>
c_scatter_offsets;
statically_indexed_array<statically_indexed_array<bool, MPerThread>, NumMEpiTile>
c_scatter_valids;
auto c_coord = dram_tile_distribution.calculate_index();
static_for<0, NumMEpiTile, 1>{}([&](auto mIter) {
static_for<0, MPerThread, 1>{}([&](auto m0) {
Expand All @@ -1262,6 +1264,7 @@ struct MoeFlatmmKernel
scatter_token_id =
scatter_token_id * kargs.TopK + (fused_token >> token_id_offset);
c_scatter_offsets[mIter][m0] = scatter_token_id * kargs.stride_C;
c_scatter_valids[mIter][m0] = (scatter_token_id < kargs.NumTokens);
});
});

Expand Down Expand Up @@ -1302,7 +1305,8 @@ struct MoeFlatmmKernel
c_block_window.get_window_lengths(),
c_block_window.get_window_origin(),
dram_tile_distribution,
c_scatter_offsets[mIter]);
c_scatter_offsets[mIter],
c_scatter_valids[mIter]);

if constexpr(!IsInputGemm ||
EpiloguePipeline::MemoryOperation == memory_operation_enum::atomic_add)
Expand Down