-
Notifications
You must be signed in to change notification settings - Fork 38
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
can akg descrip a whole network model #8
Comments
HI, mingli @MingliSun. Akg also supports composite operators which are descripted by a json file. In fact, if u enable Graph kernel fusion on MindSpore(https://www.mindspore.cn/tutorial/zh-CN/r0.5/advanced_use/graph_kernel_fusion.html), you can see the detailed compiling process of akg |
@anyrenwei Thanks for your reply,Graph kernel fusion only support ascend.Now I found all needed to do to implement a network using tvm.compute (tvm.te.compute) is defining different operators and call them sequentially.so I write a simple demo (lenet),but a check failed errors occured.here is my code and error message:
error messges:
what's that check failed mean?if my way to write a network is wrong,please give some tips to modify it.If akg don't support that for now,can you give me a demo about a typical network in cuda or ascend(especially ascend) .Using tvm.compute or topi or something else are fine.And my akg version is updated to latest release. |
First, Graph kernel fusion also support GPU. You can run Bert on MindSpore and it will enable graph kernel fusion defaultly. If you want to check all the fusion operators we now supported, you can just export the env below: |
Second, the error you listed seems that the auto-tiling module does not support the gemm format of you defined... If you want to run related operators on Ascend, I suggest you use these operators directly which defined in: https://gitee.com/mindspore/akg/tree/master/tests/common/test_op |
@anyrenwei Thanks for your reply.For now I run multiple operators in cuda(cause all I have is ascend310 which is not supported as told on another issue),and I try using topi.nn not define own operators using tvm.compute,here is my code:
And an error occurred .
so is that a akg bug?How to solve it ?And can you give me any network examples such as lenet,resnet,vgg... works fine. |
Hi,I noticed that we can pass compute/hybrid or autodiff to akg,but how to descrip a whole network model which includes lot of operator? should combine all operators into a tvm.compute,then we can can akg.build(schdule,args,...).
I'm confusing about that.
Thanks a lot.
The text was updated successfully, but these errors were encountered: