diff --git a/.github/workflows/pr-test.yml b/.github/workflows/pr-test.yml
index 894eaa061..8ff7f6eee 100644
--- a/.github/workflows/pr-test.yml
+++ b/.github/workflows/pr-test.yml
@@ -674,6 +674,10 @@ jobs:
     runs-on: 2-gpu-runner
     env:
       RUNNER_LABELS: 2-gpu-runner
+    strategy:
+      fail-fast: false
+      matrix:
+        partition: [0, 1]
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -700,7 +704,7 @@ jobs:
           if [[ "${{ needs.check-changes.outputs.continue_on_error }}" == "true" ]]; then
             CONTINUE_ON_ERROR_FLAG="--continue-on-error"
           fi
-          python3 run_suite.py --hw cuda --suite stage-b-test-large-2-gpu $CONTINUE_ON_ERROR_FLAG
+          python3 run_suite.py --hw cuda --suite stage-b-test-large-2-gpu --auto-partition-id ${{ matrix.partition }} --auto-partition-size 2 $CONTINUE_ON_ERROR_FLAG
 
   stage-c-test-large-4-gpu:
     needs: [check-changes, call-gate, stage-b-test-small-1-gpu, stage-b-test-large-1-gpu, stage-b-test-large-2-gpu, sgl-kernel-build-wheels]
@@ -1027,57 +1031,6 @@ jobs:
           IS_BLACKWELL=1 python3 run_suite.py --hw cuda --suite stage-b-test-4-gpu-b200 $CONTINUE_ON_ERROR_FLAG
 
 
-  unit-test-backend-2-gpu:
-    needs: [check-changes, call-gate, unit-test-backend-1-gpu]
-    if: |
-      always() &&
-      (
-        (inputs.target_stage == 'unit-test-backend-2-gpu') ||
-        (
-          !inputs.target_stage &&
-          (github.event_name == 'schedule' || (!failure() && !cancelled())) &&
-          ((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
-        )
-      )
-    runs-on: 2-gpu-runner
-    env:
-      RUNNER_LABELS: 2-gpu-runner
-    strategy:
-      fail-fast: false
-      matrix:
-        part: [0, 1]
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ inputs.pr_head_sha || inputs.ref || github.sha }}
-
-      - name: Download artifacts
-        if: needs.check-changes.outputs.sgl_kernel == 'true'
-        uses: actions/download-artifact@v4
-        with:
-          path: sgl-kernel/dist/
-          merge-multiple: true
-          pattern: wheel-python3.10-cuda12.9
-
-      - name: Install dependencies
-        run: |
-          CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} bash scripts/ci/ci_install_dependency.sh
-
-      - name: Run test
-        timeout-minutes: 30
-        run: |
-          cd test/srt
-          RETRY_FLAG=""
-          if [[ "${{ needs.check-changes.outputs.enable_retry }}" == "true" ]]; then
-            RETRY_FLAG="--enable-retry"
-          fi
-          CONTINUE_ON_ERROR_FLAG=""
-          if [[ "${{ needs.check-changes.outputs.continue_on_error }}" == "true" ]]; then
-            CONTINUE_ON_ERROR_FLAG="--continue-on-error"
-          fi
-          python3 run_suite.py --suite per-commit-2-gpu --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 $RETRY_FLAG $CONTINUE_ON_ERROR_FLAG
-
   unit-test-backend-4-gpu:
     needs: [check-changes, call-gate, unit-test-backend-1-gpu, stage-b-test-4-gpu-b200]
     if: |
@@ -1802,7 +1755,6 @@ jobs:
         stage-c-test-large-4-gpu,
         quantization-test,
         unit-test-backend-1-gpu,
-        unit-test-backend-2-gpu,
         stage-b-test-4-gpu-b200,
         unit-test-backend-4-gpu,
         unit-test-backend-8-gpu-h20,
diff --git a/python/sglang/multimodal_gen/runtime/loader/weight_utils.py b/python/sglang/multimodal_gen/runtime/loader/weight_utils.py
index 9210f38d8..17391c577 100644
--- a/python/sglang/multimodal_gen/runtime/loader/weight_utils.py
+++ b/python/sglang/multimodal_gen/runtime/loader/weight_utils.py
@@ -219,6 +219,26 @@ def safetensors_weights_iterator(
                     yield name, param
 
 
+def _load_pt_file(bin_file: str, device: str) -> dict:
+    """Load a PyTorch checkpoint file, handling legacy tar format.
+
+    PyTorch 2.6 changed the default of weights_only from False to True.
+    Legacy tar format files cannot be loaded with weights_only=True.
+    This function tries weights_only=True first, then falls back to False
+    for legacy tar format files from trusted sources (HuggingFace Hub).
+    """
+    try:
+        return torch.load(bin_file, map_location=device, weights_only=True)
+    except RuntimeError as e:
+        if "legacy .tar format" in str(e):
+            logger.warning(
+                "Loading %s with weights_only=False (legacy tar format)",
+                os.path.basename(bin_file),
+            )
+            return torch.load(bin_file, map_location=device, weights_only=False)
+        raise
+
+
 def pt_weights_iterator(
     hf_weights_files: list[str],
     to_cpu: bool = True,
@@ -234,7 +254,7 @@ def pt_weights_iterator(
         disable=not enable_tqdm,
         bar_format=_BAR_FORMAT,
     ):
-        state = torch.load(bin_file, map_location=device, weights_only=True)
+        state = _load_pt_file(bin_file, device)
         yield from state.items()
         del state
 
diff --git a/python/sglang/srt/distributed/device_communicators/pynccl_allocator.py b/python/sglang/srt/distributed/device_communicators/pynccl_allocator.py
index 67740409f..c678c0998 100644
--- a/python/sglang/srt/distributed/device_communicators/pynccl_allocator.py
+++ b/python/sglang/srt/distributed/device_communicators/pynccl_allocator.py
@@ -65,7 +65,10 @@ _active_symmetric_memory_context = None
 
 
 def is_symmetric_memory_enabled():
-    return get_global_server_args().enable_symm_mem
+    try:
+        return get_global_server_args().enable_symm_mem
+    except ValueError:
+        return False
 
 
 def set_graph_pool_id(graph_pool_id):
diff --git a/python/sglang/srt/model_loader/weight_utils.py b/python/sglang/srt/model_loader/weight_utils.py
index 59a9da951..f483774f2 100644
--- a/python/sglang/srt/model_loader/weight_utils.py
+++ b/python/sglang/srt/model_loader/weight_utils.py
@@ -822,6 +822,26 @@ def multi_thread_safetensors_weights_iterator(
                 yield name, param
 
 
+def _load_pt_file(bin_file: str) -> dict:
+    """Load a PyTorch checkpoint file, handling legacy tar format.
+
+    PyTorch 2.6 changed the default of weights_only from False to True.
+    Legacy tar format files cannot be loaded with weights_only=True.
+    This function tries weights_only=True first, then falls back to False
+    for legacy tar format files from trusted sources (HuggingFace Hub).
+    """
+    try:
+        return torch.load(bin_file, map_location="cpu", weights_only=True)
+    except RuntimeError as e:
+        if "legacy .tar format" in str(e):
+            logger.warning(
+                "Loading %s with weights_only=False (legacy tar format)",
+                os.path.basename(bin_file),
+            )
+            return torch.load(bin_file, map_location="cpu", weights_only=False)
+        raise
+
+
 def pt_weights_iterator(
     hf_weights_files: List[str],
 ) -> Generator[Tuple[str, torch.Tensor], None, None]:
@@ -835,7 +855,7 @@ def pt_weights_iterator(
         disable=not enable_tqdm,
         bar_format=_BAR_FORMAT,
     ):
-        state = torch.load(bin_file, map_location="cpu", weights_only=True)
+        state = _load_pt_file(bin_file)
         yield from state.items()
         del state
 
@@ -849,12 +869,9 @@ def multi_thread_pt_weights_iterator(
         not torch.distributed.is_initialized() or torch.distributed.get_rank() == 0
     )
 
-    def _load_file(bin_file: str):
-        return torch.load(bin_file, map_location="cpu", weights_only=True)
-
     with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
         futures = [
-            executor.submit(_load_file, bin_file) for bin_file in hf_weights_files
+            executor.submit(_load_pt_file, bin_file) for bin_file in hf_weights_files
         ]
 
         if enable_tqdm:
diff --git a/scripts/ci/slash_command_handler.py b/scripts/ci/slash_command_handler.py
index 14ab0c245..d1cc5f9c2 100644
--- a/scripts/ci/slash_command_handler.py
+++ b/scripts/ci/slash_command_handler.py
@@ -155,7 +155,6 @@ def handle_rerun_stage(
         "multimodal-gen-test-2-gpu",
         "quantization-test",
         "unit-test-backend-1-gpu",
-        "unit-test-backend-2-gpu",
         "stage-b-test-4-gpu-b200",
         "unit-test-backend-4-gpu",
         "unit-test-backend-8-gpu-h200",
diff --git a/test/srt/test_disaggregation_basic.py b/test/registered/disaggregation/test_disaggregation_basic.py
similarity index 99%
rename from test/srt/test_disaggregation_basic.py
rename to test/registered/disaggregation/test_disaggregation_basic.py
index c012c051c..bc2e77f1d 100644
--- a/test/srt/test_disaggregation_basic.py
+++ b/test/registered/disaggregation/test_disaggregation_basic.py
@@ -7,6 +7,7 @@ import openai
 import requests
 from transformers import AutoTokenizer
 
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
 from sglang.test.server_fixtures.disaggregation_fixture import (
     PDDisaggregationServerBase,
@@ -19,6 +20,8 @@ from sglang.test.test_utils import (
     popen_launch_pd_server,
 )
 
+register_cuda_ci(est_time=400, suite="stage-b-test-large-2-gpu")
+
 
 class TestDisaggregationAccuracy(PDDisaggregationServerBase):
     @classmethod
diff --git a/test/srt/test_data_parallelism.py b/test/registered/distributed/test_data_parallelism.py
similarity index 91%
rename from test/srt/test_data_parallelism.py
rename to test/registered/distributed/test_data_parallelism.py
index e54e3a5cf..25eba4a36 100644
--- a/test/srt/test_data_parallelism.py
+++ b/test/registered/distributed/test_data_parallelism.py
@@ -5,6 +5,7 @@ from types import SimpleNamespace
 import requests
 
 from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST,
@@ -14,6 +15,9 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
+register_cuda_ci(est_time=73, suite="stage-b-test-large-2-gpu")
+register_amd_ci(est_time=73, suite="stage-b-test-large-2-gpu-amd")
+
 
 class TestDataParallelism(CustomTestCase):
     @classmethod
diff --git a/test/srt/test_dp_attention.py b/test/registered/distributed/test_dp_attention.py
similarity index 97%
rename from test/srt/test_dp_attention.py
rename to test/registered/distributed/test_dp_attention.py
index 80bf43af4..2a3d6b608 100644
--- a/test/srt/test_dp_attention.py
+++ b/test/registered/distributed/test_dp_attention.py
@@ -5,6 +5,7 @@ import requests
 
 from sglang.srt.environ import envs
 from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
 from sglang.test.kits.radix_cache_server_kit import run_radix_attention_test
 from sglang.test.run_eval import run_eval
@@ -19,6 +20,8 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
+register_cuda_ci(est_time=350, suite="stage-b-test-large-2-gpu")
+
 
 class TestDPAttentionDP2TP2(CustomTestCase):
     @classmethod
diff --git a/test/srt/test_load_weights_from_remote_instance.py b/test/registered/distributed/test_load_weights_from_remote_instance.py
similarity index 98%
rename from test/srt/test_load_weights_from_remote_instance.py
rename to test/registered/distributed/test_load_weights_from_remote_instance.py
index db094da0c..128b9a86f 100644
--- a/test/srt/test_load_weights_from_remote_instance.py
+++ b/test/registered/distributed/test_load_weights_from_remote_instance.py
@@ -24,6 +24,7 @@ import torch
 import torch.multiprocessing as mp
 
 import sglang as sgl
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_PORT_FOR_SRT_TEST_RUNNER,
     DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
@@ -37,6 +38,9 @@ from sglang.utils import terminate_process
 
 mp.set_start_method("spawn", force=True)
 
+register_cuda_ci(est_time=72, suite="stage-b-test-large-2-gpu")
+register_amd_ci(est_time=72, suite="stage-b-test-large-2-gpu-amd")
+
 
 def verify_params_close(params1, params2, error_msg):
     """Verify if two parameter arrays are close enough."""
diff --git a/test/srt/hicache/test_hicache_storage_3fs_backend.py b/test/registered/hicache/test_hicache_storage_3fs_backend.py
similarity index 93%
rename from test/srt/hicache/test_hicache_storage_3fs_backend.py
rename to test/registered/hicache/test_hicache_storage_3fs_backend.py
index c91772881..ec28b53c3 100644
--- a/test/srt/hicache/test_hicache_storage_3fs_backend.py
+++ b/test/registered/hicache/test_hicache_storage_3fs_backend.py
@@ -1,7 +1,7 @@
 """
 Benchmark tests for HiCache Storage with 3FS backend.
 Usage:
-    python3 -m pytest test/srt/hicache/test_hicache_storage_3fs_backend.py -v
+    python3 -m pytest test/registered/hicache/test_hicache_storage_3fs_backend.py -v
 """
 
 import json
@@ -10,8 +10,11 @@ import unittest
 
 from test_hicache_storage_file_backend import HiCacheStorageBaseMixin
 
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
+register_cuda_ci(est_time=200, suite="stage-b-test-large-2-gpu")
+
 
 class HiCacheStorage3FSBackendBaseMixin(HiCacheStorageBaseMixin):
     """Base mixin class with common setup and utilities"""
diff --git a/test/srt/hicache/test_hicache_storage_file_backend.py b/test/registered/hicache/test_hicache_storage_file_backend.py
similarity index 98%
rename from test/srt/hicache/test_hicache_storage_file_backend.py
rename to test/registered/hicache/test_hicache_storage_file_backend.py
index 382db07b3..153536fc3 100644
--- a/test/srt/hicache/test_hicache_storage_file_backend.py
+++ b/test/registered/hicache/test_hicache_storage_file_backend.py
@@ -1,7 +1,7 @@
 """
 E2E tests for HiCache Storage functionality.
 Usage:
-    python3 -m pytest test/srt/hicache/test_hicache_storage_e2e.py -v
+    python3 -m pytest test/registered/hicache/test_hicache_storage_file_backend.py -v
 """
 
 import json
@@ -18,6 +18,7 @@ import requests
 
 from sglang.bench_serving import get_tokenizer
 from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
 from sglang.test.test_utils import (
     DEFAULT_MLA_MODEL_NAME_FOR_TEST,
@@ -29,6 +30,8 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
+register_cuda_ci(est_time=200, suite="stage-b-test-large-2-gpu")
+
 
 class HiCacheStorageBaseMixin:
     """Base mixin class with common setup and utilities"""
diff --git a/test/srt/hicache/test_hicache_storage_mooncake_backend.py b/test/registered/hicache/test_hicache_storage_mooncake_backend.py
similarity index 98%
rename from test/srt/hicache/test_hicache_storage_mooncake_backend.py
rename to test/registered/hicache/test_hicache_storage_mooncake_backend.py
index 7e0c83134..947bce776 100644
--- a/test/srt/hicache/test_hicache_storage_mooncake_backend.py
+++ b/test/registered/hicache/test_hicache_storage_mooncake_backend.py
@@ -1,7 +1,7 @@
 """
 Benchmark tests for HiCache Storage with Mooncake backend.
 Usage:
-    python3.10 -m pytest test/srt/hicache/test_hicache_storage_mooncake_backend.py -v
+    python3.10 -m pytest test/registered/hicache/test_hicache_storage_mooncake_backend.py -v
 """
 
 import os
@@ -12,6 +12,7 @@ import unittest
 import requests
 from test_hicache_storage_file_backend import HiCacheStorageBaseMixin
 
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_MLA_MODEL_NAME_FOR_TEST,
     CustomTestCase,
@@ -19,6 +20,8 @@ from sglang.test.test_utils import (
     is_in_ci,
 )
 
+register_cuda_ci(est_time=300, suite="stage-b-test-large-2-gpu")
+
 
 class HiCacheStorageMooncakeBackendBaseMixin(HiCacheStorageBaseMixin):
     """Base mixin class with common setup and utilities"""
diff --git a/test/srt/models/test_kimi_linear_models.py b/test/registered/models/test_kimi_linear_models.py
similarity index 91%
rename from test/srt/models/test_kimi_linear_models.py
rename to test/registered/models/test_kimi_linear_models.py
index 25e280f1f..88ef6d7ad 100644
--- a/test/srt/models/test_kimi_linear_models.py
+++ b/test/registered/models/test_kimi_linear_models.py
@@ -2,6 +2,7 @@ import unittest
 from types import SimpleNamespace
 
 from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.few_shot_gsm8k import run_eval
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -10,6 +11,8 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
+register_cuda_ci(est_time=90, suite="stage-b-test-large-2-gpu")
+
 
 class TestKimiLinear(CustomTestCase):
     @classmethod
diff --git a/test/srt/models/test_nvidia_nemotron_nano_v2.py b/test/registered/models/test_nvidia_nemotron_nano_v2.py
similarity index 96%
rename from test/srt/models/test_nvidia_nemotron_nano_v2.py
rename to test/registered/models/test_nvidia_nemotron_nano_v2.py
index 656eef838..21a5a9aa8 100644
--- a/test/srt/models/test_nvidia_nemotron_nano_v2.py
+++ b/test/registered/models/test_nvidia_nemotron_nano_v2.py
@@ -1,9 +1,12 @@
 import unittest
 
 from sglang.srt.utils import is_blackwell
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
 from sglang.test.server_fixtures.default_fixture import DefaultServerBase
 
+register_cuda_ci(est_time=132, suite="stage-b-test-large-2-gpu")
+
 
 class TestNvidiaNemotronNanoV2BF16(GSM8KMixin, DefaultServerBase):
     model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
diff --git a/test/srt/run_suite.py b/test/srt/run_suite.py
index d07dcec37..7b73c4098 100644
--- a/test/srt/run_suite.py
+++ b/test/srt/run_suite.py
@@ -16,17 +16,6 @@ suites = {
         TestFile("test_video_utils.py", 5),
         TestFile("test_modelopt_export.py", 9),
     ],
-    "per-commit-2-gpu": [
-        TestFile("hicache/test_hicache_storage_3fs_backend.py", 200),
-        TestFile("hicache/test_hicache_storage_file_backend.py", 200),
-        TestFile("hicache/test_hicache_storage_mooncake_backend.py", 300),
-        TestFile("models/test_kimi_linear_models.py", 90),
-        TestFile("models/test_nvidia_nemotron_nano_v2.py", 132),
-        TestFile("test_data_parallelism.py", 73),
-        TestFile("test_disaggregation_basic.py", 400),
-        TestFile("test_dp_attention.py", 350),
-        TestFile("test_load_weights_from_remote_instance.py", 72),
-    ],
     "per-commit-4-gpu": [
         TestFile("models/test_qwen3_next_models.py", 650),
         TestFile("test_gpt_oss_4gpu.py", 300),
@@ -116,11 +105,6 @@ suite_amd = {
         # TestFile("test_vision_chunked_prefill.py", 175), # Disabled temporarily and track in #7701
         # TestFile("test_wave_attention_backend.py", 150), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/11127
     ],
-    "per-commit-amd-mi35x": [],
-    "per-commit-2-gpu-amd": [
-        TestFile("test_data_parallelism.py", 73),
-        TestFile("test_load_weights_from_remote_instance.py", 72),
-    ],
     "per-commit-4-gpu-amd": [
         TestFile("test_pp_single_node.py", 150),
     ],