From 4e6dbe0956343d767d738547762be1aa162ef9bf Mon Sep 17 00:00:00 2001
From: LI SHENGYONG <49200266+shenchuxiaofugui@users.noreply.github.com>
Date: Fri, 20 Mar 2026 15:22:55 +0800
Subject: [PATCH] [EPLB][Bugfix] Set parallel_config.enable_eplb to true to
 load redundant experts (#7470)

### What this PR does / why we need it?
pr: https://github.com/vllm-project/vllm/pull/37136 break eplb because
it filters out redundant experts.
pr: https://github.com/vllm-project/vllm/pull/37322 fix it due to use
parallel_config.enable_eplb to determine whether to skip the weight
loading filter.
But in vllm-ascend, parallel_config.enable_eplb is always false. When we
use eplb, we temporarily set it to true.

### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->

### How was this patch tested?

![Snipaste_2026-03-19_16-13-01](https://github.com/user-attachments/assets/b3a4911e-36b3-4c31-951c-7c091f416d00)
| dataset | version | metric | mode | vllm-api-stream-chat |
|----- | ----- | ----- | ----- | -----|
| aime2024 | 604a78 | accuracy | gen | 86.67 |

Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
---
 tests/e2e/multicard/2-cards/test_qwen3_moe.py | 6 ------
 vllm_ascend/worker/model_runner_v1.py         | 5 ++++-
 2 files changed, 4 insertions(+), 7 deletions(-)

diff --git a/tests/e2e/multicard/2-cards/test_qwen3_moe.py b/tests/e2e/multicard/2-cards/test_qwen3_moe.py
index 4ce5e33e..385b32e8 100644
--- a/tests/e2e/multicard/2-cards/test_qwen3_moe.py
+++ b/tests/e2e/multicard/2-cards/test_qwen3_moe.py
@@ -76,12 +76,6 @@ def test_qwen3_moe_distributed_aiv_tp2():
 
 @pytest.mark.asyncio
 async def test_qwen3_moe_w8a8_distributed_tp2_ep_dynamic_eplb():
-    from vllm_ascend.utils import vllm_version_is
-
-    if not vllm_version_is("0.17.0"):
-        pytest.skip(
-            "EPLB output is different without EPLB, see issue: https://github.com/vllm-project/vllm-ascend/issues/7408",
-        )
     model = "vllm-ascend/Qwen3-30B-A3B-W8A8"
     port = get_open_port()
     compilation_config = json.dumps({"cudagraph_capture_sizes": [8]})
diff --git a/vllm_ascend/worker/model_runner_v1.py b/vllm_ascend/worker/model_runner_v1.py
index 1bf7095e..a49182ce 100644
--- a/vllm_ascend/worker/model_runner_v1.py
+++ b/vllm_ascend/worker/model_runner_v1.py
@@ -364,6 +364,7 @@ class NPUModelRunner(GPUModelRunner):
 
         eplb_config = self.ascend_config.eplb_config
         self.dynamic_eplb = eplb_config.dynamic_eplb
+        self.eplb_enable = self.dynamic_eplb or (eplb_config.expert_map_path is not None)
         if self.dynamic_eplb:
             self.is_eplb_warmuped = False
             self.policy_type = eplb_config.eplb_policy_type
@@ -2554,7 +2555,9 @@ class NPUModelRunner(GPUModelRunner):
         logger.info("Starting to load model %s...", self.model_config.model)
 
         with DeviceMemoryProfiler() as m:  # noqa: SIM117
-            self.model = get_model(vllm_config=self.vllm_config)
+            if self.eplb_enable:
+                self.vllm_config.parallel_config.enable_eplb = True
+            self.model: nn.Module = get_model(vllm_config=self.vllm_config)
             if self.dynamic_eplb:
                 model_register(self.model)
             if self.drafter: