#include "ggml-cuda/common.cuh"
#include "ggml.h"
#include "topk-moe.cuh"
#include <cmath>
#include <initializer_list>
// Kernel config struct - passed by value to CUDA kernel
struct topk_moe_config {
bool use_sigmoid;
bool with_norm;
bool delayed_softmax;
};
// Warp-local softmax used for both the pre-top-k logits and the post-top-k delayed path.
template <int experts_per_thread, bool use_limit>
__device__ void softmax_warp_inplace(float (&vals)[experts_per_thread], const int limit, const int lane) {
float max_val = -INFINITY;
#pragma unroll
for (int i = 0; i < experts_per_thread; i++) {
const int idx = lane + i * WARP_SIZE;
const bool active = !use_limit || (idx < limit);
if (active) {
max_val = max(max_val, vals[i]);
}
}
max_val = warp_reduce_max(max_val);
float sum = 0.f;
#pragma unroll
for (int i = 0; i < experts_per_thread; i++) {
const int idx = lane + i * WARP_SIZE;
const bool active = !use_limit || (idx < limit);
if (active) {
const float val = expf(vals[i] - max_val);
vals[i] = val;
sum += val;
} else {
vals[i] = 0.f;
}
}
sum = warp_reduce_sum(sum);
const float inv_sum = 1.0f / sum;
#pragma unroll
for (int i = 0; i < experts_per_thread; i++) {
const int idx = lane + i * WARP_SIZE;
const bool active = !use_limit || (idx < limit);
if (active) {
vals[i] *= inv_sum;
}
}
}
template <int experts_per_thread, bool use_limit>
__device__ void sigmoid_warp_inplace(float (&vals)[experts_per_thread], const int limit, const int lane) {
#pragma unroll
for (int i = 0; i < experts_per_thread; i++) {
const int idx = lane + i * WARP_SIZE;
const bool active = !use_limit || (idx < limit);
vals[i] = active ? 1.f / (1.f + expf(-vals[i])) : -INFINITY;
}
}
/*
This kernel does the following:
1. optionally softmax over the logits per token [n_experts, n_tokens]
2. argmax reduce over the top-k (n_experts_used) logits
3. write weights + ids to global memory
4. optionally normalize the weights or apply softmax over the selected logits
It is intended as fusion of softmax->top-k->get_rows pipeline for MoE models
*/
template <int n_experts, bool has_bias>
__launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * logits,
float * weights,
int32_t * ids,
float * bias,
const int n_rows,
const int n_expert_used,
const float clamp_val,
const float scale_val,
const topk_moe_config config) {
const int row = blockIdx.x * blockDim.y + threadIdx.y;
if (row >= n_rows) {
return;
}
logits += n_experts * row;
weights += n_expert_used * row;
ids += n_experts * row;
constexpr int experts_per_thread = (n_experts > WARP_SIZE) ? n_experts / WARP_SIZE : 1;
float wt[experts_per_thread];
// Initialize all slots to -INFINITY
#pragma unroll
for (int i = 0; i < experts_per_thread; i++) {
wt[i] = -INFINITY;
}
#pragma unroll
for (int i = 0; i < n_experts; i += WARP_SIZE) {
const int expert = i + threadIdx.x;
wt[i / WARP_SIZE] = (n_experts % WARP_SIZE == 0 || expert < n_experts) ? logits[expert] : -INFINITY;
}
if (!config.delayed_softmax) {
if (config.use_sigmoid) {
sigmoid_warp_inplace<experts_per_thread, false>(wt, n_experts, threadIdx.x);
} else {
softmax_warp_inplace<experts_per_thread, false>(wt, n_experts, threadIdx.x);
}
}
// Sanitize NaN to -FLT_MAX so the iterative argmax produces unique expert IDs.
// NaN comparisons always return false, which would cause the same expert to be
// selected repeatedly. -FLT_MAX compares normally and is still excluded by the
// -INFINITY sentinel used after each selection round.
// More relevant for the cuBLAS path. See https://github.com/ggml-org/llama.cpp/issues/19659
#pragma unroll
for (int i = 0; i < experts_per_thread; i++) {
if (__isnanf(wt[i])) {
wt[i] = -FLT_MAX;
}
}
// selection_wt is only needed when bias is present (selection uses wt + bias)
// when no bias, we use wt directly for both selection and weight values
float selection_wt[has_bias ? experts_per_thread : 1];
if constexpr (has_bias) {
#pragma unroll
for (int i = 0; i < experts_per_thread; i++) {
selection_wt[i] = -INFINITY;
}
#pragma unroll
for (int i = 0; i < n_experts; i += WARP_SIZE) {
const int expert = i + threadIdx.x;
selection_wt[i / WARP_SIZE] =
(n_experts % WARP_SIZE == 0 || expert < n_experts) ? wt[i / WARP_SIZE] + bias[expert] : -INFINITY;
}
}
//at this point, each thread holds either a portion of the softmax distribution
//or the raw logits. We do the argmax reduce over n_expert_used, each time marking
//the expert weight as -inf to exclude from the next iteration
float wt_sum = 0.f;
float output_weights[experts_per_thread];
#pragma unroll
for (int i = 0; i < experts_per_thread; i++) {
output_weights[i] = 0.f;
}
for (int k = 0; k < n_expert_used; k++) {
float max_val = wt[0];
int max_expert = threadIdx.x;
if constexpr (has_bias) {
float max_val_s = selection_wt[0];
#pragma unroll
for (int i = 1; i < experts_per_thread; i++) {
const int expert = threadIdx.x + i * WARP_SIZE;
if ((n_experts % WARP_SIZE == 0 || expert < n_experts) && selection_wt[i] > max_val_s) {
max_val = wt[i];
max_val_s = selection_wt[i];
max_expert = expert;
}
}
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask /= 2) {
const float val = __shfl_xor_sync(0xFFFFFFFF, max_val, mask, WARP_SIZE);
const float val_s = __shfl_xor_sync(0xFFFFFFFF, max_val_s, mask, WARP_SIZE);
const int expert = __shfl_xor_sync(0xFFFFFFFF, max_expert, mask, WARP_SIZE);
if (val_s > max_val_s || (val_s == max_val_s && expert < max_expert)) {
max_val = val;
max_val_s = val_s;
max_expert = expert;
}
}
if ((max_expert & (WARP_SIZE - 1)) == threadIdx.x) {
selection_wt[max_expert / WARP_SIZE] = -INFINITY;
}
} else {
#pragma unroll
for (int i = 1; i < experts_per_thread; i++) {
const int expert = threadIdx.x + i * WARP_SIZE;
if ((n_experts % WARP_SIZE == 0 || expert < n_experts) && wt[i] > max_val) {
max_val = wt[i];
max_expert = expert;
}
}
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask /= 2) {
const float val = __shfl_xor_sync(0xFFFFFFFF, max_val, mask, WARP_SIZE);
const int expert = __shfl_xor_sync(0xFFFFFFFF, max_expert, mask, WARP_SIZE);
if (val > max_val || (val == max_val && expert < max_expert)) {
max_val = val;
max_expert = expert;
}
}
if ((max_expert & (WARP_SIZE - 1)) == threadIdx.x) {
wt[max_expert / WARP_SIZE] = -INFINITY;
}
}
if ((k & (WARP_SIZE - 1)) == threadIdx.x) {
output_weights[k / WARP_SIZE] = max_val;
}
if ((max_expert & (WARP_SIZE - 1)) == threadIdx.x) {
ids[k] = max_expert;
if (config.with_norm) {
wt_sum += max_val;
}
}
}
if (config.with_norm) {
wt_sum = warp_reduce_sum(wt_sum);
wt_sum = max(wt_sum, clamp_val);
const float inv_sum = 1.0f / wt_sum;
for (int i = 0; i < experts_per_thread; i++) {
output_weights[i] *= inv_sum;
}
}
if (config.delayed_softmax) {
softmax_warp_inplace<experts_per_thread, true>(output_weights, n_expert_used, threadIdx.x);
}
#pragma unroll
for (int i = 0; i < experts_per_thread; i++) {
const int idx = i * WARP_SIZE + threadIdx.x;
if (idx < n_expert_used) {
weights[idx] = output_weights[i] * scale_val;
}
}
}
template<bool has_bias>
static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx,
const float * logits,
float * weights,
int32_t * ids,
float * bias,
const int n_rows,
const int n_expert,
const int n_expert_used,
const float clamp_val,
const float scale_val,
const topk_moe_config config) {
GGML_ASSERT(!(config.with_norm && config.delayed_softmax) &&
"delayed softmax is not supported with weight normalization");
const int rows_per_block = 4;
dim3 grid_dims((n_rows + rows_per_block - 1) / rows_per_block, 1, 1);
dim3 block_dims(WARP_SIZE, rows_per_block, 1);
cudaStream_t stream = ctx.stream();
switch (n_expert) {
case 1:
topk_moe_cuda<1, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
case 2:
topk_moe_cuda<2, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
case 4:
topk_moe_cuda<4, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
case 8:
topk_moe_cuda<8, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
case 16:
topk_moe_cuda<16, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
case 32:
topk_moe_cuda<32, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
case 64:
topk_moe_cuda<64, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
case 128:
topk_moe_cuda<128, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
case 256:
topk_moe_cuda<256, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
case 512:
topk_moe_cuda<512, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
case 576:
topk_moe_cuda<576, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
clamp_val, scale_val, config);
break;
default:
GGML_ASSERT(false && "fatal error");
break;
}
}
void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
const ggml_tensor * logits,
ggml_tensor * weights,
ggml_tensor * ids,
const ggml_tensor * clamp,
const ggml_tensor * scale,
const ggml_tensor * bias,
const ggml_cuda_topk_moe_args & args) {
GGML_ASSERT(logits->type == GGML_TYPE_F32);
GGML_ASSERT(weights->type == GGML_TYPE_F32);
GGML_ASSERT(ids->type == GGML_TYPE_I32);
const int n_experts = logits->ne[0];
const int n_rows = logits->ne[1];
const float * logits_d = (const float *) logits->data;
float * weights_d = (float *) weights->data;
int32_t * ids_d = (int32_t *) ids->data;
float * bias_d = bias ? (float *) bias->data : nullptr;
float scale_val = scale ? ggml_get_op_params_f32(scale, 0) : 1.0f;
GGML_ASSERT(ids->nb[1] / ggml_type_size(ids->type) == (size_t) n_experts);
const int n_expert_used = weights->ne[1];
const bool with_norm = clamp != nullptr;
float clamp_val = -INFINITY;
if (clamp) {
clamp_val = ggml_get_op_params_f32(clamp, 0);
}
topk_moe_config config;
config.use_sigmoid = args.sigmoid;
config.with_norm = with_norm;
config.delayed_softmax = args.delayed_softmax;
if (bias) {
launch_topk_moe_cuda<true>(ctx, logits_d, weights_d, ids_d, bias_d, n_rows, n_experts, n_expert_used, clamp_val,
scale_val, config);
} else {
launch_topk_moe_cuda<false>(ctx, logits_d, weights_d, ids_d, bias_d, n_rows, n_experts, n_expert_used, clamp_val,
scale_val, config);
}
}
bool ggml_cuda_should_use_topk_moe(const ggml_tensor * gating_op,
const ggml_tensor * weights,
const ggml_tensor * logits,
const ggml_tensor * ids) {
const int n_expert = ids->nb[1] / ids->nb[0];
if (((n_expert & (n_expert - 1)) != 0 || n_expert > 512) && n_expert != 576) {
return false;
}
if (!ggml_is_contiguous(weights) || !ggml_is_contiguous(logits)) {
return false;
}
if (gating_op->op == GGML_OP_SOFT_MAX) {
const ggml_tensor * softmax = gating_op;
float scale = 1.0f;
float max_bias = 0.0f;
memcpy(&scale, (const float *) softmax->op_params + 0, sizeof(float));
memcpy(&max_bias, (const float *) softmax->op_params + 1, sizeof(float));
if (!ggml_is_contiguous(softmax->src[0])) {
return false;
}
if (scale != 1.0f || max_bias != 0.0f) {
return false;
}
// don't fuse when masks or sinks are present
if (softmax->src[1] || softmax->src[2]) {
return false;
}
} else if (gating_op->op == GGML_OP_UNARY) {
ggml_unary_op op = ggml_get_unary_op(gating_op);
if (op != GGML_UNARY_OP_SIGMOID) {
return false;
}
}
return true;
}
