Re-introduce the unit test of test_mooncake_ep_small (#16019)

python/sglang/srt/eplb/eplb_algorithms/__init__.py

@@ -54,7 +54,7 @@ def rebalance_experts(
             num_groups=num_groups,
             num_nodes=num_nodes,
             num_gpus=num_physical_experts // num_local_physical_experts,
-            enable_hierarchical=True,
+            enable_hierarchical=False,
             active_ranks=(
                 ElasticEPStateManager.instance().active_ranks
                 if ElasticEPStateManager.instance() is not None

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

@@ -2,7 +2,11 @@ from __future__ import annotations
 
 import logging
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, NamedTuple, Optional
+from enum import Enum, auto
+from typing import NamedTuple, Optional
+
+import torch
+import torch.distributed as dist
 
 from sglang.srt.elastic_ep.elastic_ep import ElasticEPStateManager
 from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
@@ -18,14 +22,6 @@ from sglang.srt.layers.moe.topk import TopKOutput
 from sglang.srt.layers.moe.utils import DeepEPMode
 from sglang.srt.utils import get_int_env_var
 
-if TYPE_CHECKING:
-    from sglang.srt.batch_overlap.single_batch_overlap import CombineOverlapArgs
-
-from enum import Enum, auto
-
-import torch
-import torch.distributed as dist
-
 logger = logging.getLogger(__name__)
 
 
@@ -147,8 +143,6 @@ class _MooncakeEPDispatcherImpl:
         self.first_execution = True
         self.timeout_us = 10000000
 
-        self.active_ranks = ElasticEPStateManager.instance().active_ranks
-
         self.handle = None
 
     def dispatch_a(
@@ -215,11 +209,12 @@ class _MooncakeEPDispatcherImpl:
         use_fp8: bool = False,
     ):
         buffer = self._get_buffer()
+        active_ranks = ElasticEPStateManager.instance().active_ranks
         packed_recv_hidden, packed_recv_count, self.handle, event, hook = (
             buffer.dispatch(
                 hidden_states,
                 topk_ids,
-                self.active_ranks,
+                active_ranks,
                 self.num_max_dispatch_tokens_per_rank,
                 self.num_experts,
                 -1 if self.first_execution else self.timeout_us,
@@ -235,14 +230,13 @@ class _MooncakeEPDispatcherImpl:
         hidden_states: torch.Tensor,
         topk_ids: torch.Tensor,
         topk_weights: torch.Tensor,
-        overlap_args: Optional[CombineOverlapArgs] = None,
     ):
         hidden_states, event, hook = self._combine_core(
             hidden_states,
             topk_ids,
             topk_weights,
         )
-        return hidden_states, event, hook, overlap_args
+        return hidden_states, event, hook
 
     def combine_b(self, hidden_states, event, hook):
         hook() if self.return_recv_hook else event.current_stream_wait()
@@ -255,11 +249,12 @@ class _MooncakeEPDispatcherImpl:
         topk_weights: torch.Tensor,
     ):
         buffer = self._get_buffer()
+        active_ranks = ElasticEPStateManager.instance().active_ranks
         combined_hidden_states, event, hook = buffer.combine(
             hidden_states,
             topk_ids,
             topk_weights,
-            self.active_ranks,
+            active_ranks,
             -1 if self.first_execution else self.timeout_us,
             self.handle,
             async_finish=not self.return_recv_hook,
@@ -368,7 +363,6 @@ class MooncakeEPDispatcher(BaseDispatcher):
             hidden_states=hidden_states,
             topk_ids=topk_ids,
             topk_weights=topk_weights,
-            overlap_args=self.overlap_args,
         )
         self._combine_intermediate_state = inner_state
 

test/srt/ep/test_mooncake_ep_small.py

@@ -35,7 +35,7 @@ class TestTP(CustomTestCase):
                 "--mooncake-ib-device",
                 ib_devices,
                 "--moe-a2a-backend",
-                "deepep",
+                "mooncake",
                 "--deepep-mode",
                 "low_latency",
                 "--chunked-prefill-size",
@@ -70,6 +70,7 @@ class TestTP(CustomTestCase):
         self.assertGreater(metrics["accuracy"], 0.60)
 
 
+@unittest.skip("covered in TestMooncakeWithEPLB")
 class TestPureDP(TestTP):
     extra_args = [
         "--tp",
@@ -90,6 +91,7 @@ class TestHybridDPTP(TestTP):
     ]
 
 
+@unittest.skip("covered in TestMooncakeWithEPLB")
 class TestNoGatherdBuffer(TestTP):
     extra_args = [
         "--tp",
@@ -102,6 +104,7 @@ class TestNoGatherdBuffer(TestTP):
     ]
 
 
+@unittest.skip("covered in TestMooncakeWithEPLB")
 class TestTBO(TestTP):
     extra_args = [
         "--tp",
@@ -115,7 +118,7 @@ class TestTBO(TestTP):
     ]
 
 
-class TestMooncakeWitchEPLB(TestTP):
+class TestMooncakeWithEPLB(TestTP):
     extra_args = [
         "--tp",
         "4",

test/srt/run_suite.py

@@ -45,8 +45,7 @@ suites = {
     ],
     "per-commit-4-gpu-deepep": [
         TestFile("ep/test_deepep_small.py", 531),
-        # TODO: Add it back after mooncake supports torch 2.9
+        TestFile("ep/test_mooncake_ep_small.py", 660),
-        # TestFile("ep/test_mooncake_ep_small.py", 450),
     ],
     "per-commit-8-gpu-h200-deepep": [
         TestFile("ep/test_deepep_large.py", 563),