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

[Sync] Update to latest code (#2679)

Browse Source 
* [Sync] Update to latest code

* Add new code files

* Add new code files

* update code

* Try to fix build.sh

* Try to fix build.sh

* Update code

* Update requirements.txt

* Update code

---------

Co-authored-by: Jiang-Jia-Jun <jiangjiajun@baidu.com>
This commit is contained in:
Jiang-Jia-Jun
2025-07-03 15:43:53 +08:00
committed by GitHub
parent d222248d00
commit 05c670e593
95 changed files with 9916 additions and 1312 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
custom_ops/gpu_ops/per_token_quant_fp8.cu
View File

@@ -50,11 +50,13 @@ __global__ void quant_per_token_per_block(const T *input,
max_value_thread = max(abs(load_vec_float[vid]), max_value_thread);
}
// get max value per warp
max_value_thread = max(__shfl_xor_sync(0xffffffff, max_value_thread, 16), max_value_thread);
max_value_thread = max(__shfl_xor_sync(0xffffffff, max_value_thread, 8), max_value_thread);
max_value_thread = max(__shfl_xor_sync(0xffffffff, max_value_thread, 4), max_value_thread);
max_value_thread = max(__shfl_xor_sync(0xffffffff, max_value_thread, 2), max_value_thread);
max_value_thread = max(__shfl_xor_sync(0xffffffff, max_value_thread, 1), max_value_thread);
max_value_thread = max(__shfl_down_sync(0xffffffff, max_value_thread, 16), max_value_thread);
max_value_thread = max(__shfl_down_sync(0xffffffff, max_value_thread, 8), max_value_thread);
max_value_thread = max(__shfl_down_sync(0xffffffff, max_value_thread, 4), max_value_thread);
max_value_thread = max(__shfl_down_sync(0xffffffff, max_value_thread, 2), max_value_thread);
max_value_thread = max(__shfl_down_sync(0xffffffff, max_value_thread, 1), max_value_thread);
// broadcast max_value
max_value_thread = __shfl_sync(0xFFFFFFFF, max_value_thread, 0);
max_value_thread = max(max_value_thread, epsilon);
float scale_to_store = max_value_thread / MAX_VALUE;
// quant