Skip to content

fix(moe): initialize graph decode routes under EP #1430

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

Draft
PanZezhong1725 wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

fix(moe): initialize graph decode routes under EP #1430

Show file tree
Hide file tree
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
178 changes: 122 additions & 56 deletions src/infiniop/ops/moe_fused_dense/nvidia/moe_fused_dense_nvidia.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,7 +71,6 @@ size_t workspace_size(const MoeFusedDenseInfo &info) {
bytes += align_up(info.num_experts * sizeof(cutlass::gemm::GemmCoord), 16);
bytes += align_up(info.num_experts * sizeof(void *), 16) * 4;
bytes += align_up(info.num_experts * sizeof(int64_t), 16) * 4;
bytes += align_up(sizeof(int), 16);
return bytes + 256;
#endif
}
Expand Down Expand Up @@ -291,29 +290,59 @@ __global__ void setup_prefill_gemm2_kernel(cutlass::gemm::GemmCoord *problems,
}

template <typename T>
__global__ void setup_decode_gemm1_aligned_compact_kernel(cutlass::gemm::GemmCoord *problems,
void **ptr_a,
void **ptr_b,
void **ptr_c,
void **ptr_d,
int64_t *lda,
int64_t *ldb,
int64_t *ldc,
int64_t *ldd,
int *active_count,
int *output_permutation,
const T *hidden,
const T *w13,
T *gate_up,
const int *sorted_token_ids,
const int *expert_ids,
const int *num_tokens_post_padded,
int pairs,
int topk,
int num_experts,
int hidden_size,
int intermediate_size,
int block_size) {
__global__ void setup_decode_gemm1_defaults_kernel(cutlass::gemm::GemmCoord *problems,
void **ptr_a,
void **ptr_b,
void **ptr_c,
void **ptr_d,
int64_t *lda,
int64_t *ldb,
int64_t *ldc,
int64_t *ldd,
const T *hidden,
const T *w13,
T *gate_up,
int topk,
int hidden_size,
int intermediate_size) {
int route = blockIdx.x * blockDim.x + threadIdx.x;
if (route >= topk) {
return;
}
problems[route] = cutlass::gemm::GemmCoord(1, intermediate_size * 2, hidden_size);
ptr_a[route] = const_cast<T *>(hidden);
ptr_b[route] = const_cast<T *>(w13);
ptr_c[route] = gate_up + static_cast<size_t>(route) * intermediate_size * 2;
ptr_d[route] = gate_up + static_cast<size_t>(route) * intermediate_size * 2;
lda[route] = hidden_size;
ldb[route] = hidden_size;
ldc[route] = intermediate_size * 2;
ldd[route] = intermediate_size * 2;
}

template <typename T>
__global__ void setup_decode_gemm1_local_kernel(cutlass::gemm::GemmCoord *problems,
void **ptr_a,
void **ptr_b,
void **ptr_c,
void **ptr_d,
int64_t *lda,
int64_t *ldb,
int64_t *ldc,
int64_t *ldd,
int *output_permutation,
const T *hidden,
const T *w13,
T *gate_up,
const int *sorted_token_ids,
const int *expert_ids,
const int *num_tokens_post_padded,
int pairs,
int topk,
int num_experts,
int hidden_size,
int intermediate_size,
int block_size) {
int row = blockIdx.x * blockDim.x + threadIdx.x;
int total_rows = *num_tokens_post_padded;
if (row >= total_rows) {
Expand All @@ -328,7 +357,7 @@ __global__ void setup_decode_gemm1_aligned_compact_kernel(cutlass::gemm::GemmCoo
return;
}

int idx = atomicAdd(active_count, 1);
int idx = pair;
int token = pair / topk;
output_permutation[pair] = idx;
problems[idx] = cutlass::gemm::GemmCoord(1, intermediate_size * 2, hidden_size);
Expand All @@ -343,27 +372,58 @@ __global__ void setup_decode_gemm1_aligned_compact_kernel(cutlass::gemm::GemmCoo
}

template <typename T>
__global__ void setup_decode_gemm2_aligned_compact_kernel(cutlass::gemm::GemmCoord *problems,
void **ptr_a,
void **ptr_b,
void **ptr_c,
void **ptr_d,
int64_t *lda,
int64_t *ldb,
int64_t *ldc,
int64_t *ldd,
const T *activated,
const T *w2,
T *expert_out,
const int *sorted_token_ids,
const int *expert_ids,
const int *num_tokens_post_padded,
const int *output_permutation,
int pairs,
int num_experts,
int hidden_size,
int intermediate_size,
int block_size) {
__global__ void setup_decode_gemm2_defaults_kernel(cutlass::gemm::GemmCoord *problems,
void **ptr_a,
void **ptr_b,
void **ptr_c,
void **ptr_d,
int64_t *lda,
int64_t *ldb,
int64_t *ldc,
int64_t *ldd,
const T *activated,
const T *w2,
T *expert_out,
int topk,
int hidden_size,
int intermediate_size) {
int route = blockIdx.x * blockDim.x + threadIdx.x;
if (route >= topk) {
return;
}
problems[route] = cutlass::gemm::GemmCoord(1, hidden_size, intermediate_size);
ptr_a[route] = const_cast<T *>(activated + static_cast<size_t>(route) * intermediate_size);
ptr_b[route] = const_cast<T *>(w2);
ptr_c[route] = expert_out + static_cast<size_t>(route) * hidden_size;
ptr_d[route] = expert_out + static_cast<size_t>(route) * hidden_size;
lda[route] = intermediate_size;
ldb[route] = intermediate_size;
ldc[route] = hidden_size;
ldd[route] = hidden_size;
}

template <typename T>
__global__ void setup_decode_gemm2_local_kernel(cutlass::gemm::GemmCoord *problems,
void **ptr_a,
void **ptr_b,
void **ptr_c,
void **ptr_d,
int64_t *lda,
int64_t *ldb,
int64_t *ldc,
int64_t *ldd,
const T *activated,
const T *w2,
T *expert_out,
const int *sorted_token_ids,
const int *expert_ids,
const int *num_tokens_post_padded,
const int *output_permutation,
int pairs,
int num_experts,
int hidden_size,
int intermediate_size,
int block_size) {
int row = blockIdx.x * blockDim.x + threadIdx.x;
int total_rows = *num_tokens_post_padded;
if (row >= total_rows) {
Expand Down Expand Up @@ -521,36 +581,42 @@ infiniStatus_t calculate_typed(const MoeFusedDenseInfo &info,
auto grouped_ldb = reinterpret_cast<int64_t *>(advance_workspace(ptr, remaining, static_cast<size_t>(num_experts) * sizeof(int64_t), alignof(int64_t)));
auto grouped_ldc = reinterpret_cast<int64_t *>(advance_workspace(ptr, remaining, static_cast<size_t>(num_experts) * sizeof(int64_t), alignof(int64_t)));
auto grouped_ldd = reinterpret_cast<int64_t *>(advance_workspace(ptr, remaining, static_cast<size_t>(num_experts) * sizeof(int64_t), alignof(int64_t)));
auto active_count = reinterpret_cast<int *>(advance_workspace(ptr, remaining, sizeof(int), alignof(int)));

if (!counts || !offsets || !output_permutation || !packed_hidden || !gate_up || !activated || !expert_out || !grouped_problems || !grouped_ptr_a || !grouped_ptr_b || !grouped_ptr_c || !grouped_ptr_d || !grouped_lda || !grouped_ldb || !grouped_ldc || !grouped_ldd || !active_count) {
if (!counts || !offsets || !output_permutation || !packed_hidden || !gate_up || !activated || !expert_out || !grouped_problems || !grouped_ptr_a || !grouped_ptr_b || !grouped_ptr_c || !grouped_ptr_d || !grouped_lda || !grouped_ldb || !grouped_ldc || !grouped_ldd) {
return INFINI_STATUS_INSUFFICIENT_WORKSPACE;
}

auto w13_t = reinterpret_cast<const T *>(w13);
auto w2_t = reinterpret_cast<const T *>(w2);
if (num_tokens == 1 && topk <= num_experts) {
cudaMemsetAsync(output_permutation, 0xff, pairs * sizeof(int), stream);
cudaMemsetAsync(active_count, 0, sizeof(int), stream);
setup_decode_gemm1_aligned_compact_kernel<T><<<(max_num_tokens_padded + 255) / 256, 256, 0, stream>>>(
setup_decode_gemm1_defaults_kernel<T><<<(topk + 255) / 256, 256, 0, stream>>>(
grouped_problems, grouped_ptr_a, grouped_ptr_b, grouped_ptr_c, grouped_ptr_d,
grouped_lda, grouped_ldb, grouped_ldc, grouped_ldd, active_count, output_permutation,
grouped_lda, grouped_ldb, grouped_ldc, grouped_ldd,
reinterpret_cast<const T *>(hidden_states), w13_t, gate_up,
topk, hidden_size, intermediate_size);
setup_decode_gemm1_local_kernel<T><<<(max_num_tokens_padded + 255) / 256, 256, 0, stream>>>(
grouped_problems, grouped_ptr_a, grouped_ptr_b, grouped_ptr_c, grouped_ptr_d,
grouped_lda, grouped_ldb, grouped_ldc, grouped_ldd, output_permutation,
reinterpret_cast<const T *>(hidden_states), w13_t, gate_up,
reinterpret_cast<const int *>(sorted_token_ids),
reinterpret_cast<const int *>(expert_ids),
reinterpret_cast<const int *>(num_tokens_post_padded),
pairs, topk, num_experts, hidden_size, intermediate_size, block_size);

// Decode has one token and exactly one valid aligned row for each of
// its top-k routes, so the compact problem count is always topk.
// A D2H count copy and stream sync cannot be captured for replay.
// Keep the problem count fixed at topk for graph replay. Routes owned
// by another EP rank retain valid dummy descriptors and are ignored
// by output_permutation when local expert outputs are combined.
CHECK_STATUS(launch_cutlass_gemm_grouped_device_meta<CutlassT>(
topk, grouped_problems, grouped_ptr_a, grouped_ptr_b, grouped_ptr_c, grouped_ptr_d,
grouped_lda, grouped_ldb, grouped_ldc, grouped_ldd, stream));

swiglu_kernel<T><<<(topk * intermediate_size + 255) / 256, 256, 0, stream>>>(gate_up, activated, topk, intermediate_size);

setup_decode_gemm2_aligned_compact_kernel<T><<<(max_num_tokens_padded + 255) / 256, 256, 0, stream>>>(
setup_decode_gemm2_defaults_kernel<T><<<(topk + 255) / 256, 256, 0, stream>>>(
grouped_problems, grouped_ptr_a, grouped_ptr_b, grouped_ptr_c, grouped_ptr_d,
grouped_lda, grouped_ldb, grouped_ldc, grouped_ldd,
activated, w2_t, expert_out, topk, hidden_size, intermediate_size);
setup_decode_gemm2_local_kernel<T><<<(max_num_tokens_padded + 255) / 256, 256, 0, stream>>>(
grouped_problems, grouped_ptr_a, grouped_ptr_b, grouped_ptr_c, grouped_ptr_d,
grouped_lda, grouped_ldb, grouped_ldc, grouped_ldd, activated, w2_t, expert_out,
reinterpret_cast<const int *>(sorted_token_ids),
Expand Down
15 changes: 15 additions & 0 deletions test/infinicore/ops/moe_fused_dense.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,6 +115,8 @@ def torch_moe_fused_dense(
for token in range(num_tokens):
for route in range(topk):
expert = int(ids[token, route].item())
if expert < 0 or expert >= w13f.shape[0]:
continue
gate_up = F.linear(hidden[token], w13f[expert])
gate, up = gate_up.chunk(2, dim=-1)
activated = F.silu(gate) * up
Expand All @@ -138,6 +140,19 @@ def parse_test_cases():
block_size=4,
),
),
(
"decode EP-filtered routes",
_make_case(
20260728,
num_tokens=1,
hidden_size=32,
intermediate_size=64,
num_experts=4,
topk=4,
topk_ids=torch.tensor([[6, 1, 3, 7]], dtype=torch.int32),
block_size=4,
),
),
(
"prefill aligned missing experts",
_make_case(
Expand Down
Loading