Tiny support customize DeepEP max dispatch tokens per rank (#6934)

python/sglang/srt/layers/moe/ep_moe/token_dispatcher.py

@@ -6,7 +6,7 @@ from sglang.srt.managers.expert_distribution import (
     get_global_expert_distribution_recorder,
 )
 from sglang.srt.managers.schedule_batch import global_server_args_dict
-from sglang.srt.utils import DeepEPMode, load_json_config
+from sglang.srt.utils import DeepEPMode, get_int_env_var, load_json_config
 
 try:
     from deep_ep import Buffer, Config
@@ -188,7 +188,9 @@ class _DeepEPDispatcherImplBase:
         self.deepep_mode = deepep_mode
 
         self.params_bytes = 2
-        self.num_max_dispatch_tokens_per_rank = 128
+        self.num_max_dispatch_tokens_per_rank = get_int_env_var(
+            "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK", 128
+        )
 
         self.handle = None