EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

fix deepep assert when PD disaggregation == null (#8274)

Browse Source 
Co-authored-by: fujianhao.fjh <fujianhao.fjh@antgroup.com>
This commit is contained in:
AlphaBaby
2025-09-19 00:31:01 +08:00
committed by GitHub
parent db71c38fcd
commit 4039c626e2
1 changed files with 11 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
python/sglang/srt/layers/moe/token_dispatcher/deepep.py
View File

@@ -164,10 +164,19 @@ class DeepEPBuffer:
num_rdma_bytes, num_rdma_bytes,
) )
# We should calculate num_qps_per_rank consistently with DeepEP's test script logic:
if deepep_mode == DeepEPMode.NORMAL: if deepep_mode == DeepEPMode.NORMAL:
num_qps_per_rank = DeepEPConfig.get_instance().num_sms // 2 # refer: https://github.com/deepseek-ai/DeepEP/blob/main/tests/test_internode.py#L235
elif deepep_mode in [DeepEPMode.LOW_LATENCY, DeepEPMode.AUTO]: num_qps_per_rank = DeepEPConfig.get_instance().num_sms
elif deepep_mode == DeepEPMode.LOW_LATENCY:
# refer: https://github.com/deepseek-ai/DeepEP/blob/main/tests/test_low_latency.py#L176
num_qps_per_rank = num_experts // group.size() num_qps_per_rank = num_experts // group.size()
elif deepep_mode == DeepEPMode.AUTO:
# low-latency and normal mode all need run
# refer: https://github.com/deepseek-ai/DeepEP/blob/main/tests/test_internode.py#L235
num_qps_per_rank = max(
DeepEPConfig.get_instance().num_sms, num_experts // group.size()
)
else: else:
raise NotImplementedError raise NotImplementedError