apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[Others] add PADDLE_ENFORCE (#5288)

Browse Source
This commit is contained in:
周周周
2025-11-28 14:23:35 +08:00
committed by GitHub
parent 1539fd6056
commit 95243f012c
2 changed files with 9 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
custom_ops/gpu_ops/moe/ep_moe_expert_dispatch.cu
View File

@@ -964,7 +964,7 @@ std::vector<paddle::Tensor> EPMoeExpertDispatchFP8(
} else {
token_rows = input_dims[0];
}
const int num_rows = token_rows;
const int hidden_size = input.dims()[input_dims.size() - 1];
const int num_experts_per_rank = num_experts_per_rank_tensor.dims()[0];
@@ -988,9 +988,9 @@ std::vector<paddle::Tensor> EPMoeExpertDispatchFP8(
auto dst_weights = GetEmptyTensor(
{token_nums_feed_to_ffn}, paddle::DataType::FLOAT32, place);
auto dst_indices = GetEmptyTensor(
{num_rows, num_experts_per_rank}, paddle::DataType::INT32, place);
{token_rows, num_experts_per_rank}, paddle::DataType::INT32, place);
auto permute_indices_per_token = paddle::full(
{num_experts_per_rank, num_rows}, -1, paddle::DataType::INT32, place);
{num_experts_per_rank, token_rows}, -1, paddle::DataType::INT32, place);
auto cumsum_idx_gpu =
paddle::full({num_experts_per_rank}, 0, paddle::DataType::INT32, place);
@@ -1001,7 +1001,7 @@ std::vector<paddle::Tensor> EPMoeExpertDispatchFP8(
num_experts_per_rank_tensor,
num_experts_per_rank_padded_tensor,
moe_topk,
num_rows,
token_rows,
-1,
-1,
hidden_size,

5
custom_ops/gpu_ops/per_token_quant_fp8.cu
View File

@@ -232,6 +232,11 @@ std::vector<paddle::Tensor> PerTokenQuantPadding(paddle::Tensor &input,
auto input_dim = input.dims();
const int token_num = input_dim[0];
const int hidden_size = input_dim[1];
PADDLE_ENFORCE(block_size == 128, "now only support block_size = 128");
PADDLE_ENFORCE(hidden_size % 128 == 0,
"hidden_size must be divisible by 128");
const int hidden_size_scale = hidden_size / block_size;
auto quanted_x = GetEmptyTensor(
{token_num, hidden_size}, paddle::DataType::FLOAT8_E4M3FN, input.place());