[NPU] support npu llama2-13B export & inference #8442
Conversation
|
Thanks for your contribution! |
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #8442 +/- ##
===========================================
- Coverage 55.42% 54.29% -1.14%
===========================================
Files 617 617
Lines 96281 96339 +58
===========================================
- Hits 53366 52304 -1062
- Misses 42915 44035 +1120 ☔ View full report in Codecov by Sentry. |
ronny1996
force-pushed
the
llama2_dev
branch
3 times, most recently
from
a59da09 to
0ee4655
Compare
csrc_npu/README.md
Outdated
| @@ -0,0 +1,14 @@ | |||
| # PaddleNLP 自定义 OP | |||
There was a problem hiding this comment.
这里建议不要单独创建新的csrc目录，因为后续多硬件接入会非常多，建议直接在csrc目录创建一个npu目录
There was a problem hiding this comment.
已修改，csrc_npu -> csrc/npu
csrc_npu/README.md
Outdated
|
|
||
| # 1. 安装 PaddleCustomDevice | ||
|
|
||
| 参考 [PaddleCustomDevice NPU 安装文档](https://github.com/PaddlePaddle/PaddleCustomDevice/blob/develop/backends/npu/README_cn.md) 进行安装 |
There was a problem hiding this comment.
目前CustomDevice有NPU编译后的版本吗？
There was a problem hiding this comment.
目前没有用于高性能推理的包，推荐是自行编译
| @@ -570,7 +570,7 @@ def compute_layernorm_before_qkv(self, src, i): | |||
| return ln_out | |||
|
|
|||
| def compute_qkv_linear(self, ln_out, i): | |||
| if float(paddle.version.cuda()) < 11.6: | |||
| if paddle.version.cuda() == "False" or float(paddle.version.cuda()) < 11.6: | |||
There was a problem hiding this comment.
昆仑芯推理也是走了这个逻辑吗，如果是对昆仑芯推理有影响吗？
There was a problem hiding this comment.
没有影响，这里只影响 paddle-cpu 版本 (npu用的paddle-cpu版本) 走上面这个分支，否则 float(paddle.version.cuda()) 会报错，cpu版本的paddle.version.cuda() 返回的是字符串 False
| if predictor_args.device == "npu": | ||
| from llama.npu.export_utils import process_params | ||
|
|
||
| process_params(os.path.join(export_args.output_path, predictor_args.model_prefix)) |
There was a problem hiding this comment.
这里对NPU的模型的op的attr进行修改的原因是什么了？
There was a problem hiding this comment.
NPU高性能推理时
- 权重转置后的矩阵乘性能更好
- npu的dequant scale有特殊格式
- 这里修改避免引入硬件强相关的代码到组网
PR types
New features
PR changes
Others
Description
support npu llama2-13B export & inference