[TRTLLM-13696][test] Part2.2: Migrate CPU only tests - generic, models, disagg

tests/integration/test_lists/test-db/l0_a10.yml

@@ -19,17 +19,11 @@ l0_a10:
   - unittest/_torch/test_tensor_lru_cache.py
   - unittest/_torch/test_torch_multi_arange.py
   - unittest/utils/test_util.py
-  - unittest/utils/test_logger.py
-  - unittest/_torch/test_model_config.py
-  - unittest/models/test_quant_config_utils.py
-  - unittest/quantization/test_mode.py
   - unittest/_torch/modeling/test_modeling_mistral.py
   - unittest/_torch/modeling/test_modeling_pixtral.py
   - unittest/_torch/modeling/test_modeling_cohere2.py
-  - unittest/_torch/modeling/test_nemotron_nano_preprocessing.py
   - unittest/_torch/modeling/test_modeling_parakeet.py
   - unittest/_torch/modeling/test_modeling_radio.py
-  - unittest/_torch/modeling/test_modeling_step3p7.py
   - unittest/_torch/modeling/test_modeling_minicpmv4_6.py
   - unittest/_torch/modeling/test_multimodal_encoder_mixin.py
   - unittest/_torch/sampler/test_trtllm_sampler.py
@@ -51,29 +45,13 @@ l0_a10:
   - unittest/inputs/test_multimodal_input_processor.py
   - unittest/others/test_cache_transceiver_precheck_config.py
   - unittest/others/test_cache_transceiver_precheck_run.py
-  - unittest/others/test_convert_utils.py
-  - unittest/others/test_lm_eval.py
   - unittest/others/test_lora_manager.py
-  - unittest/others/test_lora_module_count.py
-  - unittest/others/test_time_breakdown.py
   - unittest/others/test_tracing.py
-  - unittest/disaggregated/test_disagg_openai_client.py
-  - unittest/disaggregated/test_disagg_utils.py
   - unittest/disaggregated/test_coordinator_e2e.py
   - unittest/disaggregated/test_coordinator_worker.py
-  - unittest/disaggregated/test_openai_disagg_server.py
-  - unittest/disaggregated/test_openai_disagg_service.py
-  - unittest/disaggregated/test_router.py
-  - unittest/disaggregated/test_remoteDictionary.py
   - unittest/disaggregated/test_agent_multi_backends.py
-  - unittest/disaggregated/test_messenger.py
-  - unittest/disaggregated/test_disagg_cluster_manager_worker.py
-  - unittest/disaggregated/test_cluster_storage.py
   - unittest/disaggregated/test_extractor.py
-  - unittest/disaggregated/test_peer.py
-  - unittest/disaggregated/test_cache_reuse_adapter.py
   - unittest/disaggregated/test_bounce.py
-  - unittest/disaggregated/region/test_block.py
   - unittest/disaggregated/test_mamba_transfer.py
   - unittest/tools
   - unittest/usage/test_collectors.py

tests/integration/test_lists/test-db/l0_b200.yml

@@ -181,7 +181,6 @@ l0_b200:
   - kv_cache/test_kv_cache_v2_scheduler.py::TestKVCacheV2LoRA::test_lora_chunked_prefill
   - kv_cache/test_kv_cache_v2_scheduler.py::TestKVCacheV2LoRA::test_lora_eviction
   # ------------- KV Cache Iteration Stats ---------------
-  - unittest/metrics/test_collector.py
   - kv_cache/test_kv_cache_iteration_stats.py::TestKvCacheIterationStats::test_cold_start
   - kv_cache/test_kv_cache_iteration_stats.py::TestKvCacheIterationStats::test_partial_block_reuse
   - kv_cache/test_kv_cache_iteration_stats.py::TestKvCacheIterationStats::test_full_block_reuse

tests/integration/test_lists/test-db/l0_cpu.yml

@@ -15,9 +15,16 @@ l0_cpu:
   - unittest/_torch/distributed
   - unittest/_torch/executor
   - unittest/_torch/lora
+  - unittest/_torch/memory
+  - unittest/_torch/modeling
+  - unittest/_torch/models/checkpoints
   - unittest/_torch/modules
+  - unittest/_torch/multimodal
   - unittest/_torch/ray_orchestrator/single_gpu/test_cache_transceiver_comm.py
   - unittest/_torch/speculative/hw_agnostic
+  - unittest/_torch/test_model_config.py
+  - unittest/bindings
+  - unittest/disaggregated
   - unittest/executor/test_base_worker.py ISOLATION
   - unittest/executor/test_fatal_error_health_check.py
   - unittest/executor/test_ipc.py
@@ -43,4 +50,10 @@ l0_cpu:
   - unittest/llmapi/test_request_priority.py
   - unittest/llmapi/test_serialization.py
   - unittest/llmapi/test_utils.py
-  - unittest/others/test_http_utils_fail_fast.py
+  - unittest/metrics/test_collector.py
+  - unittest/models/test_quant_config_utils.py
+  - unittest/others
+  - unittest/quantization/test_mode.py
+  - unittest/scripts
+  - unittest/tools
+  - unittest/utils/test_logger.py

tests/integration/test_lists/test-db/l0_h100.yml

@@ -21,7 +21,6 @@ l0_h100:
   - unittest/_torch/debugger
   - unittest/_torch/executor
   - unittest/_torch/lora
-  - unittest/_torch/memory
   - unittest/_torch/misc
   # ------------- modules (non-MoE) ---------------
   - unittest/_torch/modules/test_mla_helix.py
@@ -71,23 +70,10 @@ l0_h100:
   - unittest/_torch/modeling/test_modeling_qwen3_5_vl_moe.py::TestQwen3_5MoeVL::test_all
   # Dense Qwen3.5-VL is the same hybrid family (sm90+ GDN kernel), so it lands here too.
   - unittest/_torch/modeling/test_modeling_qwen3_5_vl.py::TestQwen3_5VL::test_all
-  - unittest/disaggregated/test_disagg_utils.py
-  - unittest/disaggregated/test_router.py
-  - unittest/disaggregated/test_remoteDictionary.py
   - unittest/disaggregated/test_agent_multi_backends.py
-  - unittest/disaggregated/test_messenger.py
   - unittest/disaggregated/test_extractor.py
-  - unittest/disaggregated/test_peer.py
   - unittest/disaggregated/test_bounce.py
-  - unittest/disaggregated/region/test_block.py
-  - unittest/disaggregated/region/test_aux.py
-  - unittest/disaggregated/region/test_page.py
-  - unittest/disaggregated/region/test_region.py
-  - unittest/disaggregated/test_disaggregated_params.py
-  - unittest/disaggregated/test_perf_logger.py
   - unittest/disaggregated/test_sync_transfer_timing.py
-  - unittest/disaggregated/test_rank_info.py
-  - unittest/disaggregated/test_request_id.py
   - unittest/disaggregated/test_kv_transfer.py
   - unittest/disaggregated/test_kv_transfer_mp.py
   - unittest/disaggregated/test_transceiver_bounded_polling.py
@@ -362,7 +348,6 @@ l0_h100:
   - unittest/bindings # 8 mins on H100
   - unittest/kv_cache_manager_v2_tests # 4 min
   # ------------- KV Cache Iteration Stats ---------------
-  - unittest/metrics/test_collector.py
   - kv_cache/test_kv_cache_iteration_stats.py::TestKvCacheIterationStats::test_cold_start
   - kv_cache/test_kv_cache_iteration_stats.py::TestKvCacheIterationStats::test_partial_block_reuse
   - kv_cache/test_kv_cache_iteration_stats.py::TestKvCacheIterationStats::test_full_block_reuse

tests/integration/test_lists/test-db/l0_l40s.yml

@@ -19,7 +19,6 @@ l0_l40s:
   - unittest/_torch/modeling -k "modeling_vila"
   - unittest/_torch/modeling -k "modeling_nemotron_nano_v2_vl"
   - unittest/_torch/modeling/test_modeling_llava_next.py::TestLlavaNext::test_all
-  - unittest/_torch/modeling/test_modeling_llava_next.py::test_llava_next_expand_prompt_token_ids_for_mm
   - unittest/_torch/modeling/test_modeling_qwen2_5vl.py::TestQwen2_5_VL::test_all
   - unittest/_torch/modeling/test_modeling_qwen3vl_moe.py::TestQwen3VLMoe::test_all
   - unittest/_torch/modeling/test_modeling_qwen3vl.py::TestQwen3VL::test_all

tests/unittest/_torch/executor/test_benchmark_disagg.py

@@ -33,6 +33,9 @@
 from tensorrt_llm._torch.pyexecutor.llm_request import LlmRequestState
 from tensorrt_llm._torch.pyexecutor.scheduler import ScheduledRequests
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---------------------------------------------------------------------------
 # Helpers
 # ---------------------------------------------------------------------------

tests/unittest/_torch/memory/test_gms_backend.py

@@ -41,6 +41,8 @@
     _storage_nbytes,
 )
 
+pytestmark = pytest.mark.cpu_only
+
 
 @pytest.fixture(autouse=True)
 def _stub_current_device(monkeypatch):

tests/unittest/_torch/modeling/test_gemma4_multimodal.py

@@ -931,6 +931,7 @@ def _model_available():
     return os.path.isfile(os.path.join(MODEL_26B_PATH, "config.json"))
 
 
+@pytest.mark.cpu_only
 class TestGemma4InputProcessor(unittest.TestCase):
     """Tests for Gemma4InputProcessor with real model tokenizer/processor files."""
 

tests/unittest/_torch/modeling/test_modeling_afmoe.py

@@ -19,6 +19,7 @@
 from copy import deepcopy
 from unittest.mock import Mock, patch
 
+import pytest
 import torch
 
 import tensorrt_llm
@@ -112,6 +113,7 @@ def _shutdown_kv_cache_manager(kv_cache_manager: KVCacheManager) -> None:
     kv_cache_manager.shutdown()
 
 
+@pytest.mark.cpu_only
 class TestAfmoeRegistry(unittest.TestCase):
     """Verify AfmoeForCausalLM resolves through _torch auto-model registration."""
 
@@ -128,6 +130,7 @@ def test_legacy_model_map_does_not_contain_afmoe(self):
         self.assertNotIn("AfmoeForCausalLM", MODEL_MAP)
 
 
+@pytest.mark.cpu_only
 class TestAfmoeRoutingValidation(unittest.TestCase):
     """Verify routing assumption guards."""
 
@@ -158,6 +161,7 @@ def test_model_init_rejects_invalid_routing(self):
             AfmoeForCausalLM(model_config)
 
 
+@pytest.mark.cpu_only
 class TestAfmoeWeightMapper(unittest.TestCase):
     """Verify AfmoeHfWeightMapper key transformations."""
 
@@ -264,6 +268,7 @@ def test_is_special_instance_module_for_moe(self):
         self.assertFalse(self.mapper.is_special_instance_module(mock_linear))
 
 
+@pytest.mark.cpu_only
 class TestAfmoeWeightLoading(unittest.TestCase):
     """Verify AfmoeForCausalLM applies mapper preprocessing in the real load hook."""
 

tests/unittest/_torch/modeling/test_modeling_llava_next.py

@@ -104,6 +104,7 @@ def get_scenarios(self) -> List[TestLlavaNextScenario]:
         return scenarios
 
 
+@pytest.mark.cpu_only
 def test_llava_next_expand_prompt_token_ids_for_mm():
     """Test LlavaNextInputProcessor.expand_prompt_token_ids_for_mm replaces image placeholders correctly."""
     model_path = LLAVA_NEXT_7B_CONFIG["_name_or_path"]

tests/unittest/_torch/modeling/test_modeling_nemotron_nano_v2_vl.py

@@ -64,6 +64,7 @@ def _make_minimal_nano_model_config():
     )
 
 
+@pytest.mark.cpu_only
 def test_nemotron_nano_registers_native_multimodal_epd_components():
     """Native Nano VL/Omni classes advertise MM EPD support."""
     for arch in ("NemotronH_Nano_VL_V2", "NemotronH_Nano_Omni_Reasoning_V3"):
@@ -96,6 +97,7 @@ def _assert_nano_video_handoff(handoff):
     ],
     ids=["prompt", "prompt_token_ids"],
 )
+@pytest.mark.cpu_only
 def test_nemotron_nano_epd_handoff_preserves_non_contiguous_video_runs(
     input_field, input_value, asserts_encode_not_called
 ):
@@ -149,6 +151,7 @@ def test_nemotron_nano_epd_handoff_preserves_non_contiguous_video_runs(
     ],
     ids=["normal_worker", "mm_epd_worker"],
 )
+@pytest.mark.cpu_only
 def test_nemotron_nano_multimodal_encoder_load_by_worker_role(env_value, expects_encoder):
     """Encoder load depends on whether the worker runs raw MM prefill or consumes embeddings."""
     fake_encoder = MagicMock()
@@ -186,6 +189,7 @@ def test_nemotron_nano_multimodal_encoder_load_by_worker_role(env_value, expects
         vision_encoder_cls.assert_not_called()
 
 
+@pytest.mark.cpu_only
 def test_nemotron_nano_rejects_evs_attached_video_embeddings():
     """EVS needs retained-token metadata that E/P attached embeddings do not carry."""
     model = SimpleNamespace(
@@ -277,6 +281,7 @@ def nano_llm_model():
         torch.cuda.empty_cache()
 
 
+@pytest.mark.cpu_only
 @pytest.mark.parametrize("condition", ["single", "multiple"])
 @pytest.mark.parametrize("modality", ["image", "video"])
 def test_nemotron_nano_v2_vl_input_processor(data_dict_fixture, condition, modality):
@@ -558,6 +563,7 @@ def _build_inputs(prompts_subset, media_subset):
         )
 
 
+@pytest.mark.cpu_only
 class TestEncodeMultimodalDispatch:
     def _make_mock_model(self):
         """Create a minimal mock with the attributes `_encode_multimodal` needs."""
@@ -619,6 +625,7 @@ def test_encode_multimodal_unknown_modality_raises(self):
             NemotronH_Nano_VL_V2._encode_multimodal(model, [mm_param])
 
 
+@pytest.mark.cpu_only
 class TestSoundPlaceholderInjection:
     """Test the sound placeholder token's injection points.
 
@@ -688,6 +695,7 @@ def test_three_videos_first_and_third_have_audio(self):
         )
 
 
+@pytest.mark.cpu_only
 class TestEncodeMultimodalAudioOrder:
     """Test video / audio embedding order in multi-item scenarios."""
 
@@ -807,6 +815,7 @@ def test_no_audio_concat_when_sound_encoder_is_none(self):
         model._encode_audio.assert_not_called()
 
 
+@pytest.mark.cpu_only
 class TestInterleaveVideoAudioEmbeddings:
     """Directly test `_interleave_video_audio_embeddings` with synthetic data."""
 
@@ -920,6 +929,7 @@ def test_multi_clip_audio(self):
         assert torch.equal(result, expected)
 
 
+@pytest.mark.cpu_only
 class TestEncodeAudio:
     """Numerical equivalence: batched audio vs per-input encoding.
 
@@ -1011,6 +1021,7 @@ def test_empty_input(self):
         assert NemotronH_Nano_VL_V2._encode_audio(model, []) == []
 
 
+@pytest.mark.cpu_only
 class TestEncodeMultimodalContract:
     """Verify `_encode_multimodal` conforms to the contract expected by `get_multimodal_embeddings`.
 
@@ -1098,6 +1109,7 @@ def test_empty_params_returns_empty_list(self):
         assert result == []
 
 
+@pytest.mark.cpu_only
 class TestChunkedPrefillCaching:
     """Verify that `_encode_multimodal` output is compatible with `get_multimodal_embeddings`.
 

tests/unittest/_torch/modeling/test_modeling_parakeet.py

@@ -59,6 +59,7 @@ def _make_sound_config(**overrides):
     return SimpleNamespace(**defaults)
 
 
+@pytest.mark.cpu_only
 class TestParakeetExtractor:
     def test_clip_sizes_short_audio(self):
         ext = _make_extractor()

tests/unittest/_torch/modeling/test_modeling_qwen3vl.py

@@ -798,6 +798,7 @@ def test_rot_pos_ids_matches_gpu_reference(grid_thw_list):
     torch.testing.assert_close(actual, expected, atol=0, rtol=0)
 
 
+@pytest.mark.cpu_only
 def test_rot_pos_ids_lru_cache_hit():
     """Repeated (h, w, spatial_merge_size) keys must hit the lru_cache and
     return the same underlying tensor object (no recompute)."""

tests/unittest/_torch/modeling/test_modeling_qwen3vl_preprocess.py

@@ -28,11 +28,15 @@
 from types import SimpleNamespace
 from unittest.mock import MagicMock
 
+import pytest
+
 from tensorrt_llm._torch.models.modeling_qwen3vl import (
     Qwen3VLInputProcessorBase,
     _decide_do_sample_frames,
 )
 
+pytestmark = pytest.mark.cpu_only
+
 
 def _fake_video(metadata, *, n_frames=8):
     """Stand-in for VideoData: only `.frames` (count) and `.metadata` are read."""

tests/unittest/_torch/modeling/test_modeling_step3p7.py

@@ -40,11 +40,15 @@
 import types
 import unittest
 
+import pytest
 import torch
 from parameterized import parameterized
 from transformers import PretrainedConfig
 from utils.llm_data import llm_models_root
 
+pytestmark = pytest.mark.cpu_only
+
+
 # Resolve the Step3p7 checkpoints under the shared model root (LLM_MODELS_ROOT)
 # like the other modeling tests instead of hard-coding a developer workspace
 # path. The FP8 block-scale, NVFP4, and BF16 reference checkpoints share the

tests/unittest/_torch/modeling/test_nemotron_nano_preprocessing.py

@@ -32,6 +32,8 @@
 )
 from tensorrt_llm.inputs.multimodal_data import AudioData
 
+pytestmark = pytest.mark.cpu_only
+
 
 def make_tiler(**overrides):
     """Create a DynamicResolutionImageTiler with sensible defaults."""

tests/unittest/_torch/models/checkpoints/hf/test_weight_loader.py

@@ -6,6 +6,8 @@
 from tensorrt_llm._torch.models.checkpoints.base_weight_loader import ConsumableWeightsDict
 from tensorrt_llm.mapping import Mapping
 
+pytestmark = pytest.mark.cpu_only
+
 
 class MyError(Exception):
     pass

tests/unittest/_torch/models/checkpoints/laguna/__init__.py

@@ -0,0 +1,2 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0

tests/unittest/_torch/models/checkpoints/laguna/test_laguna_weight_mapper.py

@@ -22,6 +22,8 @@
 from tensorrt_llm._torch.models.modeling_laguna import LagunaHfWeightMapper
 from tensorrt_llm._torch.modules.fused_moe.interface import MoE
 
+pytestmark = pytest.mark.cpu_only
+
 
 class _FakeQuantMode:
     def __init__(self, *, nvfp4: bool = False, fp8_block_scales: bool = False):

tests/unittest/_torch/models/checkpoints/mistral/__init__.py

@@ -0,0 +1,2 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0

tests/unittest/_torch/models/checkpoints/mistral/test_config_loader.py

@@ -13,8 +13,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import pytest
+
 from tensorrt_llm._torch.models.checkpoints.mistral.config_loader import adapt_config_dict
 
+pytestmark = pytest.mark.cpu_only
+
 
 def test_mistral_large3_vision_config_preserves_max_position_embeddings():
     config = {