[Fix] KV-cache eviction mismatch across PP ranks in DeepSeek V3/R1 (#10214)

python/sglang/srt/model_executor/model_runner.py

@@ -1260,6 +1260,16 @@ class ModelRunner:
             // self.server_args.page_size
             * self.server_args.page_size
         )
+        # different pp rank may have different num of layers, so we need to reduce the max_total_num_tokens
+        if self.pp_size > 1:
+            tensor = torch.tensor(self.max_total_num_tokens, dtype=torch.int64)
+            torch.distributed.all_reduce(
+                tensor,
+                op=torch.distributed.ReduceOp.MIN,
+                group=get_world_group().cpu_group,
+            )
+            self.max_total_num_tokens = tensor.item()
+
         # create token size for hybrid cache
         if self.is_hybrid:
             self.set_num_token_hybrid()