-
Notifications
You must be signed in to change notification settings - Fork 5.6k
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
【Hackathon 5th No.102】 move fused_embedding_eltwise_layernorm/fusion_transpose_flatten_concat/fused_fc_elementwise_layernorm to phi #57865
Conversation
你的PR提交成功,感谢你对开源项目的贡献! |
…mbedding_eltwise_op also have the bug in new IR. because the wrong memory accesss
…ss the test with new IR
auto dim_output = phi::make_ddim({batch, seq_len, hidden}); | ||
out->set_dims(dim_output); | ||
// out->share_lod(ids); | ||
// context->ShareLoD("Ids", /*->*/ "Out"); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
后面可以单独提个PR 这里注释给删掉,out->share_lod(ids); 这行为什么给注释掉?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
因为在phi 中的share_lod需要一个tensor类型的变量作为参数,但是在这个算子里面的ids是一个vector<DesneTensor*>的类型,暂时不知道怎么处理,所以先留着。这种情况应该怎么处理呢@yuanlehome
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
因为在phi 中的share_lod需要一个tensor类型的变量作为参数,但是在这个算子里面的ids是一个vector<DesneTensor*>的类型,暂时不知道怎么处理,所以先留着。这种情况应该怎么处理呢@yuanlehome
share第0个,out->share_lod(*ids[0]);
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Done,等内存访问异常BUG解决后一起合并
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM for OP-Benchmark
问下,这个 “开启FLAGS_enable_new_ir_in_executor=1,单测也可以运行成功” 验证了吗? |
额, 其他两个算子的验证都通过了。 |
可以 set GLOG_v=5 看下日志,定位下出错位置 |
我在kernel中使用了VLOG的方式进行输出定位,但是在我使用命令 |
对应的实现是这个kernel paddle/phi/kernels/fusion/gpu/fused_embedding_eltwise_layernorm_kernel.cu |
好吧,我看错了。。。sorry |
…transpose_flatten_concat/fused_fc_elementwise_layernorm to phi (PaddlePaddle#57865) * transplant fused_embedding_elt_wise_layer_norm_kernel * fix the error * fix some bug * move the transpose to phi but new IR have a bug in output==nullptr. embedding_eltwise_op also have the bug in new IR. because the wrong memory accesss * remove some useless code * move fused_fc_elementwise_layernorm to phi, but have a bug in making * fix the bug in build the fused_fc_elementwise_layernorm_kernel and pass the test with new IR * try to fix the bug
…transpose_flatten_concat/fused_fc_elementwise_layernorm to phi (PaddlePaddle#57865) * transplant fused_embedding_elt_wise_layer_norm_kernel * fix the error * fix some bug * move the transpose to phi but new IR have a bug in output==nullptr. embedding_eltwise_op also have the bug in new IR. because the wrong memory accesss * remove some useless code * move fused_fc_elementwise_layernorm to phi, but have a bug in making * fix the bug in build the fused_fc_elementwise_layernorm_kernel and pass the test with new IR * try to fix the bug
这个PR #58115 修复了 test/ir/test_ir_embedding_eltwise_layernorm_fuse_pass.py 单测在 export FLAGS_enable_new_ir_in_executor=1 时的运行报错。 |
…transpose_flatten_concat/fused_fc_elementwise_layernorm to phi (PaddlePaddle#57865) * transplant fused_embedding_elt_wise_layer_norm_kernel * fix the error * fix some bug * move the transpose to phi but new IR have a bug in output==nullptr. embedding_eltwise_op also have the bug in new IR. because the wrong memory accesss * remove some useless code * move fused_fc_elementwise_layernorm to phi, but have a bug in making * fix the bug in build the fused_fc_elementwise_layernorm_kernel and pass the test with new IR * try to fix the bug
PR types
Others
PR changes
Others
Description
move fused_embedding_eltwise_layernorm/fusion_transpose_flatten_concat/fused_fc_elementwise_layernorm to phi
#57262