Support (1 <= dp < tp) in the dp attention in DeepEP (#4770)

Co-authored-by: Cheng Wan <cwan39@gatech.edu>
2025-03-27 20:09:35 -04:00
parent 98a2cfa9b2
commit 7f19e083c1
10 changed files with 238 additions and 47 deletions
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -1186,7 +1186,7 @@ class Scheduler(
                ret = None

        # Handle DP attention
-        if self.server_args.enable_dp_attention:
+        if self.server_args.enable_dp_attention or self.server_args.enable_sp_layernorm:
            ret, _ = self.prepare_dp_attn_batch(ret)

        return ret