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

[Quantization] Support w4afp8 MoE dynamic quantization (#5282)

Browse Source 
* support dynamic activation quant for w4afp8

* support dynamic w4afp8

* add test

* fix

* fix

---------

Co-authored-by: zhoutianzi666 <17801055074@163.com>
This commit is contained in:
Sunny-bot1
2025-12-02 18:56:16 +08:00
committed by GitHub
parent 429dd2b1db
commit 3629db4129
9 changed files with 478 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
custom_ops/gpu_ops/moe/ep_moe_expert_dispatch.cu
View File

@@ -430,7 +430,9 @@ __global__ void permute_x_kernel(
}
abs_max = phi::BlockAllReduce<MaxOp, float, Kthread>(abs_max);
float scale = 440.f / abs_max; // use 440 so we do not have to clip
dequant_scale[dst_token_idx] = abs_max;
if (tid == 0) {
dequant_scale[dst_token_idx] = abs_max;
}
for (int v_id = tid; v_id < hidden_size_int4; v_id += blockDim.x) {
Load<T, vec_size>(&data_smem[v_id * vec_size], &src_vec);
#pragma unroll
@@ -661,7 +663,7 @@ std::vector<paddle::Tensor> EPMoeExpertDispatch(
int dequant_scale_size = 1;
if (moe_quant_type == "w4afp8" && !up_gate_proj_in_scale) {
dequant_scale_size = moe_topk * num_rows;
dequant_scale_size = token_nums_this_rank;
}
auto dequant_scale =