[PD] Fix Assertion failed: /DeepEP/csrc/kernels/internode.cu:483, condition: ibgda_get_state()->num_rc_per_pe >= num_channels #134 (#5830)

2025-04-30 10:38:54 +08:00
parent 3c8a52311a
commit cc4a80caf6
1 changed files with 1 additions and 3 deletions
--- a/python/sglang/srt/layers/moe/ep_moe/token_dispatcher.py
+++ b/python/sglang/srt/layers/moe/ep_moe/token_dispatcher.py
@@ -84,9 +84,7 @@ class DeepEPBuffer:
            num_nvl_bytes,
            num_rdma_bytes,
            low_latency_mode=deepep_mode.enable_low_latency(),
-            num_qps_per_rank=(
-                num_experts // group.size() if deepep_mode.enable_low_latency() else 1
-            ),
+            num_qps_per_rank=(max(num_experts // group.size(), Buffer.num_sms // 2)),
        )
        return cls._buffer