Files
xc-llm-ascend/.github
sharonyunyun 941269a6c5 adjusting the communication method in graph mode (#1194)
### What this PR does / why we need it?
Communication performance optimization: replace allreduce with
reduce_scatter+all_gather in MLA layer's TP group,to remove
stridedsliced and all_gather in MOE layer.
when tp > 1, It is enabled during the decode phase of the graph mode
when enable_multistream_moe、MLA, use_v1, and MC2 are used.
According to the end-to-end RL inference test results, this PR can bring
3% gain in the decode stage.

**Before Improvement**
Profiling kernel_details

![image](https://github.com/user-attachments/assets/1bb5dfa1-809b-410a-90c9-c5fd23cff003)
Evaluation

![image](https://github.com/user-attachments/assets/0b8ea0c7-88e7-410f-9ef4-f0cfe910cdc7)

![image](https://github.com/user-attachments/assets/94fde910-c125-4c2e-8de4-88fc3fafc057)

**After Improvement**
Profiling kernel_details

![image](https://github.com/user-attachments/assets/55fac0e0-11f2-4654-8fd4-287949e0b29e)
Evaluation

![image](https://github.com/user-attachments/assets/e923f74b-29c4-4171-9382-40a00cf05df0)

![image](https://github.com/user-attachments/assets/5dba7967-07ea-4926-a8be-804bfd34e3e4)

### Does this PR introduce _any_ user-facing change?
Users need to configure enable_multistream_moe=True

### How was this patch tested?
Add e2e test cases to cover code logic

Signed-off-by: sharonyunyun <zhangying134@huawei.com>
2025-06-25 19:56:49 +08:00
..
2025-06-25 19:28:26 +08:00