bug(mocked_model/MockedMegatron): EP dispatch sizes should multiply by `topK`

(as discussed in Ding Chat with @MXtremist)

EP dispatch size,

https://github.com/aliyun/aicb/blob/9ac2f0ecc233996e8da3f34b2ae4c15f81543134/workload_generator/mocked_model/MockedMegatron.py#L431-L442

should be,

```python
topk * seq_len * hidden_size * batch_size // tp_size 
```



	workloads.append(
	LogItem(
	comm_type=CommType.all_to_all,
	comm_group=CommGroup.ep_group,
	msg_size=self.seq_len
	* self.hidden_size
	* self.batch_size
	// self.tp_size
	* 2,
	stage=f"{stage}.MoE",
	)
	)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

bug(mocked_model/MockedMegatron): EP dispatch sizes should multiply by `topK` #46

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

bug(mocked_model/MockedMegatron): EP dispatch sizes should multiply by topK #46

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

bug(mocked_model/MockedMegatron): EP dispatch sizes should multiply by `topK` #46