From d64dd3e18e4fd76b0256c01f83531a4972987694 Mon Sep 17 00:00:00 2001
From: Baizhou Zhang <sobereddiezhang@gmail.com>
Date: Sun, 16 Nov 2025 15:54:17 -0800
Subject: [PATCH] [Tiny]Fix 1-gpu nightly test bugs (#13389)

---
 .../layers/attention/nsa/test_nsa_indexer.py    |  1 +
 test/srt/lora/test_lora_qwen3_vl.py             | 17 ++---------------
 test/srt/run_suite.py                           |  5 +----
 3 files changed, 4 insertions(+), 19 deletions(-)

diff --git a/test/srt/layers/attention/nsa/test_nsa_indexer.py b/test/srt/layers/attention/nsa/test_nsa_indexer.py
index 43cd6262d..a6860d9e7 100644
--- a/test/srt/layers/attention/nsa/test_nsa_indexer.py
+++ b/test/srt/layers/attention/nsa/test_nsa_indexer.py
@@ -239,6 +239,7 @@ class TestNSAIndexer(CustomTestCase):
         }
         params.update(kwargs)
 
+        torch.set_default_dtype(self.dtype)
         indexer = Indexer(**params)
         # Move indexer to CUDA device
         indexer = indexer.to(device=self.device)
diff --git a/test/srt/lora/test_lora_qwen3_vl.py b/test/srt/lora/test_lora_qwen3_vl.py
index de64cbb33..063873640 100644
--- a/test/srt/lora/test_lora_qwen3_vl.py
+++ b/test/srt/lora/test_lora_qwen3_vl.py
@@ -1,4 +1,3 @@
-import os
 import random
 import unittest
 from typing import Sequence
@@ -8,7 +7,7 @@ from utils import TORCH_DTYPES, LoRAAdaptor, LoRAModelCase, ensure_reproducibili
 from sglang.srt.models.qwen3_vl import Qwen3VLForConditionalGeneration
 from sglang.srt.models.qwen3_vl_moe import Qwen3VLMoeForConditionalGeneration
 from sglang.test.runners import HFRunner, SRTRunner
-from sglang.test.test_utils import CustomTestCase, calculate_rouge_l, is_in_ci
+from sglang.test.test_utils import CustomTestCase, calculate_rouge_l
 
 
 class TestLoRAQwen3VLGating(CustomTestCase):
@@ -165,6 +164,7 @@ def _run_lora_multiple_batch_on_model_cases(
                 lora_backend=backend,
                 sleep_on_idle=True,
                 attention_backend="torch_native",
+                disable_radix_cache=True,
             )
 
             ensure_reproducibility()
@@ -224,19 +224,6 @@ class TestLoRAQwen3VLIntegration(CustomTestCase):
                     variant_label=label,
                 )
 
-    def test_all_lora_models(self):
-        if is_in_ci():
-            return
-        for label, model_case in LORA_MODEL_VARIANTS:
-            if "ONLY_RUN" in os.environ and os.environ["ONLY_RUN"] != model_case.base:
-                continue
-            with self.subTest(variant=label):
-                _run_lora_multiple_batch_on_model_cases(
-                    [model_case],
-                    max_new_tokens=LORA_MAX_NEW_TOKENS,
-                    variant_label=label,
-                )
-
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/test/srt/run_suite.py b/test/srt/run_suite.py
index 2752886e8..5788fbf22 100644
--- a/test/srt/run_suite.py
+++ b/test/srt/run_suite.py
@@ -209,7 +209,6 @@ suites = {
     "nightly-1-gpu": [
         TestFile("layers/attention/nsa/test_nsa_indexer.py", 2),
         TestFile("lora/test_lora_qwen3.py", 97),
-        TestFile("lora/test_lora_qwen3_vl.py", 200),
         TestFile("lora/test_lora_radix_cache.py", 200),
         TestFile("lora/test_lora_eviction_policy.py", 200),
         TestFile("lora/test_lora_openai_api.py", 30),
@@ -256,9 +255,7 @@ suites = {
         TestFile("lora/test_chunked_sgmv_backend.py"),
         TestFile("lora/test_lora_llama4.py"),
         TestFile("lora/test_lora_cuda_graph.py"),
-        TestFile("lora/test_lora_qwen3.py"),
-        TestFile("lora/test_lora_tp.py"),
-        TestFile("lora/test_lora_update.py"),
+        TestFile("lora/test_lora_qwen3_vl.py"),
         TestFile("models/test_clip_models.py"),
         TestFile("models/test_dummy_grok_models.py"),
         TestFile("models/test_falcon_h1_models.py"),