A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 23:30 UTC
Optimize moe_align_block_size for deepseek_v3 (#12850)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Feb 13, 2025 at 18:43 UTC
2344192a55022d84ba3bb8d33b1ec38724f54fed
2 files changed
+39
-16
csrc/moe/moe_align_sum_kernels.cu
+37
-15
@@ -198,26 +198,27 @@ __global__ void moe_align_block_size_global_mem_kernel(
198
}
199
200
// taken from
201
-// https://github.com/sgl-project/sglang/commit/ded9fcd09a43d5e7d5bb31a2bc3e9fc21bf65d2a
201
+// https://github.com/sgl-project/sglang/commit/cdae77b03dfc6fec3863630550b45bbfc789f957
202
template <typename scalar_t>
203
__global__ void sgl_moe_align_block_size_kernel(
204
scalar_t* __restrict__ topk_ids, int32_t* sorted_token_ids,
205
int32_t* expert_ids, int32_t* total_tokens_post_pad, int32_t num_experts,
206
int32_t block_size, size_t numel, int32_t* cumsum) {
207
__shared__ int32_t shared_counts[32][8];
208
- __shared__ int32_t local_offsets[256];
208
209
const int warp_id = threadIdx.x / 32;
211
- const int lane_id = threadIdx.x % 32;
210
const int experts_per_warp = 8;
211
const int my_expert_start = warp_id * experts_per_warp;
212
213
+ // Initialize shared_counts for this warp's experts
214
for (int i = 0; i < experts_per_warp; ++i) {
215
if (my_expert_start + i < num_experts) {
216
shared_counts[warp_id][i] = 0;
217
}
218
}
219
220
+ __syncthreads();
221
+
222
const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
223
const size_t start_idx = threadIdx.x * tokens_per_thread;
224
@@ -230,6 +231,7 @@ __global__ void sgl_moe_align_block_size_kernel(
231
232
__syncthreads();
233
234
+ // Single thread computes cumulative sum and total tokens
235
if (threadIdx.x == 0) {
236
cumsum[0] = 0;
237
for (int i = 1; i <= num_experts; ++i) {
@@ -246,19 +248,28 @@ __global__ void sgl_moe_align_block_size_kernel(
248
249
__syncthreads();
250
251
+ // Assign expert IDs to blocks
252
if (threadIdx.x < num_experts) {
253
for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1];
254
i += block_size) {
255
expert_ids[i / block_size] = threadIdx.x;
256
}
254
- local_offsets[threadIdx.x] = cumsum[threadIdx.x];
257
}
258
+}
259
257
- __syncthreads();
258
-
259
- for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
260
+// taken from
261
+// https://github.com/sgl-project/sglang/commit/cdae77b03dfc6fec3863630550b45bbfc789f957
262
+template <typename scalar_t>
263
+__global__ void sgl_moe_token_sort_kernel(scalar_t* __restrict__ topk_ids,
264
+ int32_t* sorted_token_ids,
265
+ int32_t* cumsum_buffer,
266
+ size_t numel) {
267
+ const size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
268
+ const size_t stride = blockDim.x * gridDim.x;
269
+
270
+ for (size_t i = tid; i < numel; i += stride) {
271
int32_t expert_id = topk_ids[i];
261
- int32_t rank_post_pad = atomicAdd(&local_offsets[expert_id], 1);
272
+ int32_t rank_post_pad = atomicAdd(&cumsum_buffer[expert_id], 1);
273
sorted_token_ids[rank_post_pad] = i;
274
}
275
}
@@ -377,23 +388,34 @@ void sgl_moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
388
torch::Tensor experts_ids,
389
torch::Tensor num_tokens_post_pad) {
390
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
391
+ TORCH_CHECK(num_experts == 256,
392
+ "sgl_moe_align_block_size kernel only supports deepseek v3.");
393
+
394
VLLM_DISPATCH_INTEGRAL_TYPES(
395
topk_ids.scalar_type(), "sgl_moe_align_block_size_kernel", [&] {
382
- // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
383
- // tensors
396
+ // calc needed amount of shared mem for `cumsum` tensors
397
auto options_int =
398
torch::TensorOptions().dtype(torch::kInt).device(topk_ids.device());
386
- // torch::Tensor token_cnts_buffer =
387
- // torch::empty({(num_experts + 1) * num_experts}, options_int);
399
torch::Tensor cumsum_buffer =
389
- torch::empty({num_experts + 1}, options_int);
400
+ torch::zeros({num_experts + 1}, options_int);
401
391
- auto kernel = vllm::moe::sgl_moe_align_block_size_kernel<scalar_t>;
392
- kernel<<<1, 1024, 0, stream>>>(
402
+ auto align_kernel =
403
+ vllm::moe::sgl_moe_align_block_size_kernel<scalar_t>;
404
+ align_kernel<<<1, 1024, 0, stream>>>(
405
topk_ids.data_ptr<scalar_t>(), sorted_token_ids.data_ptr<int32_t>(),
406
experts_ids.data_ptr<int32_t>(),
407
num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
408
topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>());
409
+
410
+ const int block_threads = 256;
411
+ const int num_blocks =
412
+ (topk_ids.numel() + block_threads - 1) / block_threads;
413
+ const int max_blocks = 65535;
414
+ const int actual_blocks = std::min(num_blocks, max_blocks);
415
+ auto sort_kernel = vllm::moe::sgl_moe_token_sort_kernel<scalar_t>;
416
+ sort_kernel<<<actual_blocks, block_threads, 0, stream>>>(
417
+ topk_ids.data_ptr<scalar_t>(), sorted_token_ids.data_ptr<int32_t>(),
418
+ cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel());
419
});
420
}
421
vllm/model_executor/layers/fused_moe/fused_moe.py
+2
-1
@@ -596,7 +596,7 @@ def moe_align_block_size(
596
dtype=torch.int32,
597
device=topk_ids.device)
598
if num_experts >= 224:
599
- if envs.VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON:
599
+ if envs.VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON or num_experts != 256:
600
moe_align_block_size_triton(
601
topk_ids,
602
num_experts,
@@ -606,6 +606,7 @@ def moe_align_block_size(
606
num_tokens_post_pad,
607
)
608
else:
609
+ # Currently requires num_experts=256
610
ops.sgl_moe_align_block_size(
611
topk_ids,
612
num_experts,