From 943c0b94ea7d131c4c78def0d687bb10f00fe57f Mon Sep 17 00:00:00 2001
From: fluctlux <38945811+fluctlux@users.noreply.github.com>
Date: Fri, 5 Dec 2025 23:03:10 +0800
Subject: [PATCH 1/2] fix ngram & suffix ci oom issue

Signed-off-by: fluctlux <38945811+fluctlux@users.noreply.github.com>
---
 .../spec_decode_v1/test_v1_spec_decode.py     | 19 ++++++++++++-------
 1 file changed, 12 insertions(+), 7 deletions(-)

diff --git a/tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py b/tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
index 0902fe6dd68..2d12506085c 100644
--- a/tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
+++ b/tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 from __future__ import annotations
 
+import os
 import random
 from typing import Any
 
@@ -9,6 +10,8 @@
 
 from tests.e2e.conftest import VllmRunner
 
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
 
 @pytest.fixture
 def test_prompts():
@@ -61,7 +64,6 @@ def eagle3_model_name():
     return "vllm-ascend/EAGLE3-LLaMA3.1-Instruct-8B"
 
 
-@pytest.mark.skip("TODO: Revert me after ngram oom issue on ci is fixed")
 def test_ngram_correctness(
     test_prompts: list[list[dict[str, Any]]],
     sampling_config: SamplingParams,
@@ -71,9 +73,11 @@ def test_ngram_correctness(
     Compare the outputs of a original LLM and a speculative LLM
     should be the same when using ngram speculative decoding.
     '''
-    ref_llm = LLM(model=model_name, max_model_len=1024, enforce_eager=False)
-    ref_outputs = ref_llm.chat(test_prompts, sampling_config)
-    del ref_llm
+
+    with VllmRunner(model_name, max_model_len=1024,
+                    enforce_eager=False) as ref_llm:
+        ref_outputs = ref_llm.model.chat(test_prompts, sampling_config)
+
     with VllmRunner(model_name,
                     speculative_config={
                         "method": "ngram",
@@ -156,9 +160,10 @@ def test_suffix_correctness(
     Compare the outputs of a original LLM and a speculative LLM
     should be the same when using ngram speculative decoding.
     '''
-    ref_llm = LLM(model=model_name, max_model_len=1024, enforce_eager=False)
-    ref_outputs = ref_llm.chat(test_prompts, sampling_config)
-    del ref_llm
+    with VllmRunner(model_name, max_model_len=1024,
+                    enforce_eager=False) as ref_llm:
+        ref_outputs = ref_llm.model.chat(test_prompts, sampling_config)
+
     with VllmRunner(model_name,
                     speculative_config={
                         "method": "suffix",

From ad3444f95483eaf681f02332bd6e10adef9eba59 Mon Sep 17 00:00:00 2001
From: fluctlux <38945811+fluctlux@users.noreply.github.com>
Date: Sat, 6 Dec 2025 17:21:16 +0800
Subject: [PATCH 2/2] enable e2e/test_v1_spec_decode

Signed-off-by: fluctlux <38945811+fluctlux@users.noreply.github.com>
---
 .github/workflows/_e2e_test.yaml | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/.github/workflows/_e2e_test.yaml b/.github/workflows/_e2e_test.yaml
index cab9ca927bd..5ec5fbb3d00 100644
--- a/.github/workflows/_e2e_test.yaml
+++ b/.github/workflows/_e2e_test.yaml
@@ -107,8 +107,7 @@ jobs:
           # ------------------------------------ v1 spec decode test ------------------------------------ #
           pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
           pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_torchair_correctness.py
-          # Fix me: test_eagle_correctness OOM error
-          #pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
+          pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
 
   e2e-2-cards:
     name: multicard-2