Skip to content

[TRTLLM-8637][feat] Optimize the routing kernel for DeepseekV3 (MoE CUTLASS backend); Add support for 384 experts (MoE TRTLLM backend) #7761

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
ChristinaZ merged 1 commit into from
Oct 20, 2025
Merged

[TRTLLM-8637][feat] Optimize the routing kernel for DeepseekV3 (MoE CUTLASS backend); Add support for 384 experts (MoE TRTLLM backend) #7761

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
88 changes: 84 additions & 4 deletions cpp/tensorrt_llm/kernels/moeTopKFuncs.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,7 +51,7 @@ struct TopKRedType
static __host__ __device__ inline TypeCmp makeCmpVal(T val, int32_t idx = 0)
{
auto valueBits = cub::Traits<T>::TwiddleIn(reinterpret_cast<typename cub::Traits<T>::UnsignedBits&>(val));
TypeCmp compactTmp = reinterpret_cast<TypeCmp&>(valueBits);
TypeCmp compactTmp = valueBits;
compactTmp = (compactTmp << kMoveBits) | (0xFFFF & (kMaxIdx - idx));
// Use 65535 minus idx to give higher priority to elements with smaller indices.
return compactTmp;
Expand Down Expand Up @@ -162,9 +162,28 @@ struct Sort<4, RedType>
}
};

template <int K, typename Type>
__forceinline__ __device__ void reduceTopK(cg::thread_block_tile<kWARP_SIZE> const& warp, Type (&out)[K],
int32_t (&outIdx)[K], Type value, int32_t idx, Type const minValue, int actualK = K)
{
static_assert(K > 0, "Top K must have K > 0");
static_assert(K < kWARP_SIZE, "Top K must have K < kWARP_SIZE");
using RedType = TopKRedType<Type>;
RedType topK{value, idx};
typename RedType::TypeCmp packedMax{};
#pragma unroll
for (int kk = 0; kk < actualK; ++kk) //@todo: check if actualK is correct
{
topK = kk > 0 && packedMax == topK.compValIdx ? RedType{minValue, idx} : topK;
// get the next largest value
packedMax = topK.reduce(warp);
RedType::unpack(out[kk], outIdx[kk], packedMax);
}
};

template <int K, typename Type, int N, bool IsSorted = false>
__device__ void reduceTopK(cg::thread_block_tile<kWARP_SIZE> const& warp, Type (&out)[K], int32_t (&outIdx)[K],
Type (&value)[N], int32_t (&idx)[N], Type minValue)
__device__ void reduceTopKFunc(cg::thread_block_tile<kWARP_SIZE> const& warp, Type (&out)[K], int32_t (&outIdx)[K],
Type (&value)[N], int32_t (&idx)[N], Type minValue, int actualK = K)
{
static_assert(K > 0, "Top K must have K > 0");
static_assert(K < kWARP_SIZE, "Top K must have K < kWARP_SIZE");
Expand All @@ -184,7 +203,7 @@ __device__ void reduceTopK(cg::thread_block_tile<kWARP_SIZE> const& warp, Type (
}
typename RedType::TypeCmp packedMax{};
#pragma unroll
for (int kk = 0; kk < K; ++kk)
for (int kk = 0; kk < actualK; ++kk)
{
bool update = kk > 0 && packedMax == topK[0].compValIdx;
#pragma unroll
Expand All @@ -198,6 +217,67 @@ __device__ void reduceTopK(cg::thread_block_tile<kWARP_SIZE> const& warp, Type (
}
};

template <int K, typename Type, int N>
__forceinline__ __device__ void reduceTopK(cg::thread_block_tile<kWARP_SIZE> const& warp, Type (&out)[K],
int32_t (&outIdx)[K], Type (&value)[N], int32_t (&idx)[N], Type const minValue, int actualK = K)
{
static_assert(K > 0, "Top K must have K > 0");
static_assert(K < kWARP_SIZE, "Top K must have K < kWARP_SIZE");
static_assert(N > 0, "Top K must have N > 0");
static_assert(N <= 16, "Only support candidates number less than or equal to 16*32=512");
static_assert(
N <= 4 || N % 4 == 0, "Only support candidates number is a multiple of 4*32=128 or less than or equal to 4");
using RedType = TopKRedType<Type>;

if constexpr (N <= 4)
{
reduceTopKFunc<K, Type, N>(warp, out, outIdx, value, idx, minValue, actualK);
}
else
{

constexpr int numLoops = N / 4;
constexpr int numResults = (numLoops * K - 1) / kWARP_SIZE + 1;

Type topKBufferValue[numResults];
int32_t topKBufferIdx[numResults];
int32_t laneIdx = threadIdx.x % kWARP_SIZE;

for (int ii = 0; ii < numResults; ++ii)
{
topKBufferValue[ii] = minValue;
topKBufferIdx[ii] = ii * kWARP_SIZE - 1; //@todo: check if this is correct
}
for (int loop = 0; loop < numLoops; ++loop)
{
int start = loop * 4;
Type topKValue[K];
int32_t topKIdx[K];
Type inValue[4];
int32_t inIdx[4];
for (int i = 0; i < 4; ++i)
{
inValue[i] = value[start + i];
inIdx[i] = idx[start + i];
}
reduceTopKFunc<K, Type, 4>(warp, topKValue, topKIdx, inValue, inIdx, minValue, actualK);
int inOffset = laneIdx % K;
if (laneIdx >= loop * K && laneIdx < (loop + 1) * K)
{
topKBufferValue[0] = topKValue[inOffset];
topKBufferIdx[0] = topKIdx[inOffset];
}
if (loop == numLoops - 1 && (laneIdx < (numLoops * K - kWARP_SIZE)))
{
topKBufferValue[1] = topKValue[inOffset];
topKBufferIdx[1] = topKIdx[inOffset];
}
}

reduceTopKFunc<K, Type, numResults>(warp, out, outIdx, topKBufferValue, topKBufferIdx, minValue, actualK);
}
Comment thread
ChristinaZ marked this conversation as resolved.
};

#undef TOPK_SWAP

} // namespace reduce_topk
Expand Down
Loading