Huang Zhengjie

Thanks for your suggestion. It's a good idea to have a PglNodePropPredDataset. We are going to refactor our citation_benchmark into NodePropBenchmark.

metapath2vec的训练建议使用分布式CPU模式。 我们已经将MAG240M 训练好的 m2v_embedding 提供了云盘下载的地址。具体看下面链接里面，m2v embedding部分 [https://github.com/PaddlePaddle/PGL/tree/main/examples/kddcup2021/MAG240M/r_unimp](url)

我们用MPI集群训练的，为了追求速度20台机一起训，因为训这套向量的时间已经比较久了，花的时间记录没了。你也可以按照 [Graph4Rec](https://github.com/PaddlePaddle/PGL/tree/main/apps/Graph4Rec)里面提供的配置去训，只要机器内存加起来能超过向量所需的内存，应该能跑起来了。

我们已经将MAG240M 训练好的 m2v_embedding 提供了云盘下载的地址。具体看下面链接里面，m2v embedding部分 [https://github.com/PaddlePaddle/PGL/tree/main/examples/kddcup2021/MAG240M/r_unimp](https://github.com/PaddlePaddle/PGL/issues/url)

> @usuyama brought up another good suggestion here: #3129: the ability add a single message from the bot/user side. I would have expected something like this to work: > >...

@Palang2014 Thanks a lot! Your code do great helps! And Great work @ry

https://github.com/PaddlePaddle/PGL/blob/main/examples/deploy_gnn/convert_to_static.py 建议你可以用这种方法来，动静转换，不建议直接用 graph作为forward对象。最好把边和特征显式列出来再转

因为paddle里面的jit只支持tensor输入，不支持复杂自定义数据结构。所以如果你要直接用jit.to_static应该还是要改一下的，我估计应该像下面这种改法，仿照我上面说的，估计是可以的。 ``` @paddle.jit.to_static def forward(self, num_nodes, edges, x): graph = pgl.Graph(num_nodes, edges, node_feat={"x": x} blablabla ```

如果你要搞静态图，你可以看看这里的写法 https://github.com/PaddlePaddle/PGL/blob/main/examples/deploy_gnn/convert_to_static.py ``` num_nodes = static.data(name='num_nodes', shape=[-1], dtype='int32') edges = static.data(name='edges', shape=[-1, 2], dtype='int32') feature = static.data( name="feature", shape=[-1, graph_obj["node_feat"].shape[-1]], dtype="float32") ``` 先定义一些placehodler，然后输入到网络里，在用静态图组网 ``` class GNNModel(nn.Layer): def __init__(self, input_size, output_size,...

> Thank you for the clarification! I do have a follow-up question. Since the SFT loss is computed only on the assistant tokens, I'm wondering whether it might make more...