diff --git a/tests/integration/test_lists/test-db/l0_a10.yml b/tests/integration/test_lists/test-db/l0_a10.yml
index b3b8bdb0e9fa..7c0d6ddcfe48 100644
--- a/tests/integration/test_lists/test-db/l0_a10.yml
+++ b/tests/integration/test_lists/test-db/l0_a10.yml
@@ -28,14 +28,10 @@ l0_a10:
   - unittest/_torch/modeling/test_multimodal_encoder_mixin.py
   - unittest/_torch/sampler/test_trtllm_sampler.py
   - unittest/_torch/sampler/test_token_ban.py
-  - unittest/_torch/executor/test_kv_cache_estimation.py
-  - unittest/_torch/executor/test_kv_cache_budget_split.py
-  - unittest/_torch/executor/test_kv_pool_rebalance.py
   - unittest/_torch/executor/test_disagg_index_mapper_early_release.py
   - unittest/_torch/executor/test_kv_cache_compression_manager.py
   - unittest/_torch/executor/test_kv_cache_v2_capacity_only.py
   - unittest/_torch/executor/test_error_classification.py
-  - unittest/_torch/modules/dwdp/test_dwdp_fixup_moe_backends.py
   - unittest/_torch/modules/moe/test_communication_factory.py
   # NOTE: this is a CPU-only test, but we do not have a dedicated job for this (and therefore no
   # test list either).
@@ -54,12 +50,6 @@ l0_a10:
   - unittest/disaggregated/test_bounce.py
   - unittest/disaggregated/test_mamba_transfer.py
   - unittest/tools
-  - unittest/usage/test_collectors.py
-  - unittest/usage/test_config.py
-  - unittest/usage/test_llmapi_config_telemetry_docs.py
-  - unittest/usage/test_opt_out.py
-  - unittest/usage/test_reporter.py
-  - unittest/usage/test_schema.py
   - unittest/usage/test_transport.py
   - unittest/usage/test_e2e_capture.py
   - disaggregated/test_disaggregated.py::test_disaggregated_single_gpu[TinyLlama-1.1B-Chat-v1.0]
@@ -104,12 +94,8 @@ l0_a10:
   - test_e2e.py::test_trtllm_bench_invalid_token_pytorch[TinyLlama-1.1B-Chat-v1.0-TinyLlama-1.1B-Chat-v1.0]
   # visual_gen
   - unittest/_torch/visual_gen/test_profiler.py
-  - unittest/_torch/visual_gen/test_visual_gen_params.py
-  - unittest/visual_gen/test_output.py
   - unittest/visual_gen/test_iteration_stats.py
   - unittest/visual_gen/test_media_encoding.py
-  - unittest/_torch/visual_gen/test_tensor_payload.py
-  - unittest/_torch/visual_gen/test_flux_infer.py
   - unittest/_torch/visual_gen/test_qwen_image_infer.py
   - unittest/_torch/visual_gen/test_qwen_image_pipeline.py
   # llmapi
@@ -151,7 +137,6 @@ l0_a10:
   # helper-script unit tests (CPU-only, ~3s)
   - unittest/scripts
   - unittest/kv_cache_manager_v2_tests # 4 min
-  - unittest/dynamo
   - unittest/api_stability
   - unittest/bindings
 - condition:
diff --git a/tests/integration/test_lists/test-db/l0_b200.yml b/tests/integration/test_lists/test-db/l0_b200.yml
index 9662d6a9947f..d0fb994e99e9 100644
--- a/tests/integration/test_lists/test-db/l0_b200.yml
+++ b/tests/integration/test_lists/test-db/l0_b200.yml
@@ -111,7 +111,6 @@ l0_b200:
   # ------------- MoE components tests ---------------
   - unittest/_torch/modules/test_moe_load_balancer.py
   - unittest/_torch/modules/test_moe_routing.py
-  - unittest/_torch/modules/test_moe_host_sharer.py
   - unittest/_torch/modules/fused_moe/test_deepgemm_fused_gather_finalize.py
   - unittest/_torch/modules/fused_moe/test_deepgemm_fused_expand_quant.py
   # ------------- MoE: test_moe_backend (by backend) ---------------
@@ -195,11 +194,7 @@ l0_b200:
   - kv_cache/test_prefix_aware_scheduling.py::TestServePrefixAwareScheduling::test_multi_round_qa_shared_prefix_smoke
   # ------------- Visual Gen tests ---------------
   - unittest/_torch/visual_gen/test_media_decode.py
-  - unittest/_torch/visual_gen/test_visual_gen_args.py
-  - unittest/_torch/visual_gen/test_visual_gen_params.py
-  - unittest/_torch/visual_gen/test_visual_gen_utils.py
   - unittest/_torch/visual_gen/test_warmup.py
-  - unittest/_torch/visual_gen/test_teacache.py
   - unittest/_torch/visual_gen/test_cache_dit.py
   - unittest/_torch/visual_gen/test_quant_ops.py
   - unittest/_torch/visual_gen/test_attention_cute_dsl.py
@@ -207,7 +202,6 @@ l0_b200:
   - unittest/_torch/visual_gen/test_attention_trtllm_sage.py
   - unittest/_torch/visual_gen/test_attention_integration.py
   - unittest/_torch/visual_gen/test_attention_perf.py
-  - unittest/_torch/visual_gen/test_trtllm_serve_endpoints.py
   - unittest/_torch/visual_gen/test_trtllm_serve_e2e.py
   - unittest/_torch/visual_gen/test_model_loader.py
   - unittest/_torch/visual_gen/test_flux_transformer.py
diff --git a/tests/integration/test_lists/test-db/l0_cpu.yml b/tests/integration/test_lists/test-db/l0_cpu.yml
index bf67c89236da..e4e5d9465a05 100644
--- a/tests/integration/test_lists/test-db/l0_cpu.yml
+++ b/tests/integration/test_lists/test-db/l0_cpu.yml
@@ -12,6 +12,7 @@ l0_cpu:
       backend: generic
       orchestrator: mpi
   tests:
+  - unittest/_torch/attention
   - unittest/_torch/cute_dsl/test_kimi_k3_kda_ptx_patch.py
   - unittest/_torch/distributed
   - unittest/_torch/executor
@@ -24,9 +25,26 @@ l0_cpu:
   - unittest/_torch/ray_orchestrator/single_gpu/test_cache_transceiver_comm.py
   - unittest/_torch/speculative/hw_agnostic
   - unittest/_torch/test_model_config.py
-  - unittest/_torch/thop/parallel_hw_agnostic/test_custom_ops.py
+  - unittest/_torch/thop/parallel_hw_agnostic
+  - unittest/_torch/visual_gen/multi_gpu/test_attn2d_attention.py
+  - unittest/_torch/visual_gen/multi_gpu/test_visual_gen_mapping.py
+  - unittest/_torch/visual_gen/multi_gpu/test_visual_gen_multinode.py
+  - unittest/_torch/visual_gen/test_attention_integration.py
+  - unittest/_torch/visual_gen/test_cache_dit.py
+  - unittest/_torch/visual_gen/test_flux_infer.py
+  - unittest/_torch/visual_gen/test_ltx2_pipeline.py
+  - unittest/_torch/visual_gen/test_ltx2_transformer.py
+  - unittest/_torch/visual_gen/test_teacache.py
+  - unittest/_torch/visual_gen/test_tensor_payload.py
+  - unittest/_torch/visual_gen/test_trtllm_serve_endpoints.py
+  - unittest/_torch/visual_gen/test_visual_gen_args.py
+  - unittest/_torch/visual_gen/test_visual_gen_params.py
+  - unittest/_torch/visual_gen/test_visual_gen_utils.py
+  - unittest/_torch/weight_sharing
+  - unittest/api_stability
   - unittest/bindings
   - unittest/disaggregated
+  - unittest/dynamo
   - unittest/executor/test_base_worker.py ISOLATION
   - unittest/executor/test_fatal_error_health_check.py
   - unittest/executor/test_ipc.py
@@ -35,6 +53,7 @@ l0_cpu:
   - unittest/executor/test_event_loop_error_broadcast.py
   - unittest/executor/test_stats_serializer.py
   - unittest/inputs
+  - unittest/kv_cache_manager_v2_tests
   - unittest/llmapi/apps/test_chat_utils.py
   - unittest/llmapi/apps/test_harmony_channel_validation.py
   - unittest/llmapi/apps/test_tool_parsers.py
@@ -57,6 +76,9 @@ l0_cpu:
   - unittest/models/test_quant_config_utils.py
   - unittest/others
   - unittest/quantization/test_mode.py
+  - unittest/scaffolding
   - unittest/scripts
   - unittest/tools
+  - unittest/usage
   - unittest/utils/test_logger.py
+  - unittest/visual_gen/test_output.py
diff --git a/tests/integration/test_lists/test-db/l0_dgx_b200.yml b/tests/integration/test_lists/test-db/l0_dgx_b200.yml
index 92267dafe500..2445af72e4e3 100644
--- a/tests/integration/test_lists/test-db/l0_dgx_b200.yml
+++ b/tests/integration/test_lists/test-db/l0_dgx_b200.yml
@@ -219,7 +219,6 @@ l0_dgx_b200:
   - unittest/_torch/visual_gen/multi_gpu/test_ulysses_async.py
   - unittest/_torch/visual_gen/multi_gpu/test_ulysses_sage_attention.py
   - unittest/_torch/visual_gen/multi_gpu/test_visual_gen_mapping.py
-  - unittest/_torch/visual_gen/multi_gpu/test_visual_gen_multinode.py
   - unittest/_torch/visual_gen/multi_gpu/test_wan_async_ulysses.py
   - unittest/_torch/visual_gen/multi_gpu/test_wan_pipeline_parallel.py
   - unittest/_torch/visual_gen/multi_gpu/test_wan_tp.py
diff --git a/tests/integration/test_lists/test-db/l0_dgx_b300.yml b/tests/integration/test_lists/test-db/l0_dgx_b300.yml
index 736319d8409f..03e7d9f29ed8 100644
--- a/tests/integration/test_lists/test-db/l0_dgx_b300.yml
+++ b/tests/integration/test_lists/test-db/l0_dgx_b300.yml
@@ -93,7 +93,6 @@ l0_dgx_b300:
       backend: pytorch
   tests:
   # ------------- MoE components tests (multi-GPU) ---------------
-  - unittest/_torch/modules/test_moe_host_sharer.py
   # ------------- MoE: multi-GPU module tests (DEP parallel, per backend per quant) ---------------
   # CUTLASS backend: FP8, NVFP4, W4A8_MXFP4_MXFP8, W8A16
   - unittest/_torch/modules/moe/test_moe_module.py::test_configurable_moe_multi_gpu[parallel=DEP-comm=DEEPEP-e60_k4_h2048_i1408-seq=8-dtype=torch.bfloat16-backend=CUTLASS-quant=FP8-routing=Renormalize]
diff --git a/tests/integration/test_lists/test-db/l0_gb300_multi_gpus.yml b/tests/integration/test_lists/test-db/l0_gb300_multi_gpus.yml
index 273aa9bc72ad..efe49cc75d67 100644
--- a/tests/integration/test_lists/test-db/l0_gb300_multi_gpus.yml
+++ b/tests/integration/test_lists/test-db/l0_gb300_multi_gpus.yml
@@ -32,7 +32,6 @@ l0_gb300_multi_gpus:
   # ------------- MoE components tests ---------------
   - unittest/_torch/modules/test_moe_load_balancer.py
   - unittest/_torch/modules/test_moe_routing.py
-  - unittest/_torch/modules/test_moe_host_sharer.py
   - unittest/_torch/modeling -k "modeling_llama"
   - unittest/_torch/modeling -k "modeling_mixtral"
   - unittest/_torch/modeling -k "modeling_gpt_oss"
diff --git a/tests/integration/test_lists/test-db/l0_h100.yml b/tests/integration/test_lists/test-db/l0_h100.yml
index fb57eb7c1a8f..9fc436b33c02 100644
--- a/tests/integration/test_lists/test-db/l0_h100.yml
+++ b/tests/integration/test_lists/test-db/l0_h100.yml
@@ -35,7 +35,6 @@ l0_h100:
   # ------------- MoE components tests ---------------
   - unittest/_torch/modules/test_moe_load_balancer.py
   - unittest/_torch/modules/test_moe_routing.py
-  - unittest/_torch/modules/test_moe_host_sharer.py
   - unittest/_torch/modules/fused_moe/test_deepgemm_fused_gather_finalize.py
   - unittest/_torch/modules/fused_moe/test_deepgemm_fused_expand_quant.py
   # ------------- MoE: test_moe_backend (by backend) ---------------
@@ -103,11 +102,6 @@ l0_h100:
   - unittest/others/test_kv_cache_transceiver.py::test_kv_cache_transceiver_single_process[PYTHON-mla-ctx_fp16_gen_fp16]
   - unittest/llmapi/test_llm_telemetry.py
   - unittest/scaffolding
-  - unittest/usage/test_collectors.py
-  - unittest/usage/test_config.py
-  - unittest/usage/test_opt_out.py
-  - unittest/usage/test_reporter.py
-  - unittest/usage/test_schema.py
   - unittest/usage/test_transport.py
   - unittest/usage/test_e2e_capture.py
   - accuracy/test_llm_api_pytorch.py::TestGemma3_1BInstruct::test_auto_dtype
diff --git a/tests/integration/test_lists/test-db/l0_rtx_pro_6000.yml b/tests/integration/test_lists/test-db/l0_rtx_pro_6000.yml
index f7372a8c4224..bacee43d87ec 100644
--- a/tests/integration/test_lists/test-db/l0_rtx_pro_6000.yml
+++ b/tests/integration/test_lists/test-db/l0_rtx_pro_6000.yml
@@ -15,7 +15,6 @@ l0_rtx_pro_6000:
   tests:
   # ------------- PyTorch tests ---------------
   - unittest/_torch/modeling -k "modeling_out_of_tree"
-  - unittest/_torch/modules/moe/test_cute_dsl_b12x_moe_backend.py
   # - unittest/_torch/modeling -k "modeling_qwen" # https://nvbugs/5234573
   - unittest/_torch/attention/test_attention_mla.py
   # SM120 W4A16 / W4A8 mixed-dtype GEMM coverage (paired with FinegrainedMixedDtypeGemm
diff --git a/tests/unittest/_torch/attention/sparse/dsa/test_dsa_indexer.py b/tests/unittest/_torch/attention/sparse/dsa/test_dsa_indexer.py
index 66578c3d7707..b322004baef1 100644
--- a/tests/unittest/_torch/attention/sparse/dsa/test_dsa_indexer.py
+++ b/tests/unittest/_torch/attention/sparse/dsa/test_dsa_indexer.py
@@ -2136,6 +2136,7 @@ def test_compute_cu_seqlen_bounds_with_cache_compressed_kv():
     ],
     ids=["small_single", "small_multi", "large_chunked", "mixed", "exact_size", "non_zero_start"],
 )
+@pytest.mark.cpu_only
 def test_split_prefill_chunks(max_chunk_size, seq_lens, start_idx, expected_specs):
     """
     Test split_prefill_chunks covering:
@@ -3278,6 +3279,7 @@ def test_indexer_topk_multi_request_with_different_cache(enable_indexer_skip):
         )
 
 
+@pytest.mark.cpu_only
 class TestPrepareRestoreAttnMetadataForDraftReplay:
     """Tests for prepare_attn_metadata_for_draft_replay and
     restore_attn_metadata_after_draft_replay."""
diff --git a/tests/unittest/_torch/attention/test_attention_mla.py b/tests/unittest/_torch/attention/test_attention_mla.py
index c8e0690593e0..5c1d66b903ac 100644
--- a/tests/unittest/_torch/attention/test_attention_mla.py
+++ b/tests/unittest/_torch/attention/test_attention_mla.py
@@ -362,10 +362,12 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
 num_generation_steps = [10]
 
 # tokens_per_block = 32 for blackwell
-tokens_per_block = 32 if torch.cuda.get_device_capability() >= (10, 0) else 64
+cuda_capability = torch.cuda.get_device_capability() if torch.cuda.device_count(
+) > 0 else (0, 0)
+tokens_per_block = 32 if cuda_capability >= (10, 0) else 64
 
 kv_cache_dtype_list = [torch.bfloat16]
-if torch.cuda.get_device_capability() in [(8, 9), (9, 0), (10, 0), (12, 0)]:
+if cuda_capability in [(8, 9), (9, 0), (10, 0), (12, 0)]:
     kv_cache_dtype_list.append(torch.float8_e4m3fn)
 scenarios = [
     Scenario(kv_cache_dtype=kv_cache_dtype,
@@ -388,6 +390,7 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
         (100, "chunked_prefill"),
     ],
 )
+@pytest.mark.cpu_only
 def test_mla_chunked_prefill_dispatch_by_sm(sm_version, expected_path,
                                             monkeypatch):
     import tensorrt_llm._torch.modules.mla as mla_module
@@ -544,8 +547,7 @@ def test_attention_mla_flashinfer(scenario: Scenario,
                                   v2_kv_cache: bool):
     """Test FlashInfer MLA computation for both context and generation phases"""
     pytest.importorskip("flashinfer")
-    if (not torch.cuda.is_available()
-            or torch.cuda.get_device_capability() != (10, 0)):
+    if (not torch.cuda.is_available() or cuda_capability != (10, 0)):
         pytest.skip("FlashInfer MLA test only runs on SM100 (Blackwell)")
 
     num_heads = scenario.num_heads
@@ -868,8 +870,8 @@ def yarn_get_mscale(scale=1, mscale=1):
                 mapping=mapping,
             )
             if backend_name == "TRTLLM":
-                gen_metadata_kwargs["enable_flash_mla"] = (
-                    torch.cuda.get_device_capability() == (9, 0))
+                gen_metadata_kwargs["enable_flash_mla"] = (cuda_capability == (
+                    9, 0))
             attn_metadata = AttentionCls.Metadata(**gen_metadata_kwargs)
             attn_metadata.prepare()
         for layer_idx in range(num_layers):
diff --git a/tests/unittest/_torch/attention/test_attention_op_sync.py b/tests/unittest/_torch/attention/test_attention_op_sync.py
index cfd56978b848..d1fe94562f55 100644
--- a/tests/unittest/_torch/attention/test_attention_op_sync.py
+++ b/tests/unittest/_torch/attention/test_attention_op_sync.py
@@ -42,6 +42,8 @@
 import typing
 from dataclasses import fields
 
+import pytest
+
 from tensorrt_llm._torch.attention_backend.fmha.fallback import (
     _THOP_EXCLUDED_FIELDS,
     _THOP_LITERALS,
@@ -50,6 +52,9 @@
 from tensorrt_llm._torch.attention_backend.interface import AttentionForwardArgs
 from tensorrt_llm._torch.attention_backend.trtllm import TrtllmAttention, TrtllmAttentionMetadata
 
+pytestmark = pytest.mark.cpu_only
+
+
 # Roots used as the LHS of attribute chains at the call site. Match the
 # names inside ``FallbackFmha.forward``.
 _SOURCE_CLASSES = {
diff --git a/tests/unittest/_torch/executor/test_dual_pool_kv_cache.py b/tests/unittest/_torch/executor/test_dual_pool_kv_cache.py
index 06e0231a1a0b..5c86f87d8df1 100644
--- a/tests/unittest/_torch/executor/test_dual_pool_kv_cache.py
+++ b/tests/unittest/_torch/executor/test_dual_pool_kv_cache.py
@@ -35,6 +35,9 @@
 from tensorrt_llm.llmapi.llm_args import CapacitySchedulerPolicy, KvCacheConfig, TorchLlmArgs
 from tensorrt_llm.mapping import Mapping
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---------------------------------------------------------------------------
 # Helpers
 # ---------------------------------------------------------------------------
diff --git a/tests/unittest/_torch/executor/test_kv_cache_budget_split.py b/tests/unittest/_torch/executor/test_kv_cache_budget_split.py
index 109460a5a57b..1cf44999e676 100644
--- a/tests/unittest/_torch/executor/test_kv_cache_budget_split.py
+++ b/tests/unittest/_torch/executor/test_kv_cache_budget_split.py
@@ -21,6 +21,9 @@
 from tensorrt_llm._torch.pyexecutor._util import CacheCost, KvCacheCreator
 from tensorrt_llm.llmapi.llm_args import KvCacheConfig
 
+pytestmark = pytest.mark.cpu_only
+
+
 GB = 1 << 30
 
 
diff --git a/tests/unittest/_torch/executor/test_kv_cache_estimation.py b/tests/unittest/_torch/executor/test_kv_cache_estimation.py
index 5c75c9186bd7..dacc6ac64efe 100644
--- a/tests/unittest/_torch/executor/test_kv_cache_estimation.py
+++ b/tests/unittest/_torch/executor/test_kv_cache_estimation.py
@@ -21,6 +21,9 @@
 from tensorrt_llm._torch.pyexecutor.kv_cache_manager_v2 import KVCacheManagerV2
 from tensorrt_llm.llmapi.llm_args import KvCacheConfig, MultimodalConfig
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---------------------------------------------------------------------------
 # Helpers
 # ---------------------------------------------------------------------------
diff --git a/tests/unittest/_torch/executor/test_kv_cache_v2_multimodal_runs.py b/tests/unittest/_torch/executor/test_kv_cache_v2_multimodal_runs.py
index 6bb5c46d2b31..6995d556398c 100644
--- a/tests/unittest/_torch/executor/test_kv_cache_v2_multimodal_runs.py
+++ b/tests/unittest/_torch/executor/test_kv_cache_v2_multimodal_runs.py
@@ -11,6 +11,9 @@
     gen_multimodal_cache_key_tokens,
 )
 
+pytestmark = pytest.mark.cpu_only
+
+
 _HASH_INTS = (1, 2, 3, 4, 5, 6, 7, 8)
 _OTHER_HASH_INTS = (8, 7, 6, 5, 4, 3, 2, 1)
 
diff --git a/tests/unittest/_torch/executor/test_kv_cache_v2_scheduler.py b/tests/unittest/_torch/executor/test_kv_cache_v2_scheduler.py
index 0a7c67e9e702..22dc0638d01c 100644
--- a/tests/unittest/_torch/executor/test_kv_cache_v2_scheduler.py
+++ b/tests/unittest/_torch/executor/test_kv_cache_v2_scheduler.py
@@ -25,6 +25,9 @@
 from tensorrt_llm._torch.pyexecutor.llm_request import LlmRequestState
 from tensorrt_llm.llmapi.llm_args import CapacitySchedulerPolicy, ContextChunkingPolicy
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---------------------------------------------------------------------------
 # State value constants
 # ---------------------------------------------------------------------------
diff --git a/tests/unittest/_torch/executor/test_kv_pool_rebalance.py b/tests/unittest/_torch/executor/test_kv_pool_rebalance.py
index ed0cc5636e7b..13ea630b3a9a 100644
--- a/tests/unittest/_torch/executor/test_kv_pool_rebalance.py
+++ b/tests/unittest/_torch/executor/test_kv_pool_rebalance.py
@@ -34,6 +34,9 @@
 from tensorrt_llm._torch.pyexecutor.py_executor import PyExecutor
 from tensorrt_llm.runtime.kv_cache_manager_v2 import OutOfPagesError
 
+pytestmark = pytest.mark.cpu_only
+
+
 # --------------------------------------------------------------------------- #
 # Helpers
 # --------------------------------------------------------------------------- #
diff --git a/tests/unittest/_torch/executor/test_kvcache_aware_router.py b/tests/unittest/_torch/executor/test_kvcache_aware_router.py
index 900151c7ac1d..1f173ed0da83 100644
--- a/tests/unittest/_torch/executor/test_kvcache_aware_router.py
+++ b/tests/unittest/_torch/executor/test_kvcache_aware_router.py
@@ -19,12 +19,17 @@
 
 from unittest.mock import MagicMock, Mock
 
+import pytest
+
 from tensorrt_llm._torch.pyexecutor.scheduler.adp_router import (
     ADPRouter,
     KVCacheAwareADPRouter,
     RankState,
 )
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---- Helpers ----
 
 
diff --git a/tests/unittest/_torch/modules/dwdp/test_dwdp_fixup_moe_backends.py b/tests/unittest/_torch/modules/dwdp/test_dwdp_fixup_moe_backends.py
index d5392e6890c8..a845ab192455 100644
--- a/tests/unittest/_torch/modules/dwdp/test_dwdp_fixup_moe_backends.py
+++ b/tests/unittest/_torch/modules/dwdp/test_dwdp_fixup_moe_backends.py
@@ -31,6 +31,7 @@
 import unittest
 from unittest.mock import MagicMock
 
+import pytest
 import torch
 import torch.nn as nn
 
@@ -40,6 +41,8 @@
     _scatter_shards_to_full,
 )
 
+pytestmark = pytest.mark.cpu_only
+
 
 def _make_mock_comm(all_shards):
     """Create a mock MPI comm whose allgather() returns ``all_shards``.
diff --git a/tests/unittest/_torch/modules/moe/test_cute_dsl_b12x_moe_backend.py b/tests/unittest/_torch/modules/moe/test_cute_dsl_b12x_moe_backend.py
index 1feac564b176..b98bb563a947 100644
--- a/tests/unittest/_torch/modules/moe/test_cute_dsl_b12x_moe_backend.py
+++ b/tests/unittest/_torch/modules/moe/test_cute_dsl_b12x_moe_backend.py
@@ -35,6 +35,9 @@
 from tensorrt_llm._torch.modules.fused_moe.fused_moe_cutlass import CutlassFusedMoE
 from tensorrt_llm.models.modeling_utils import QuantAlgo, QuantConfig
 
+pytestmark = pytest.mark.cpu_only
+
+
 _FUSED_MOE_MODULE = "tensorrt_llm._torch.modules.fused_moe.fused_moe_cute_dsl_b12x"
 
 
diff --git a/tests/unittest/_torch/modules/test_moe_host_sharer.py b/tests/unittest/_torch/modules/test_moe_host_sharer.py
index 0ed0ee609bb8..a725572373e2 100644
--- a/tests/unittest/_torch/modules/test_moe_host_sharer.py
+++ b/tests/unittest/_torch/modules/test_moe_host_sharer.py
@@ -1,12 +1,15 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 from mpi4py import MPI
 
 from tensorrt_llm._torch.modules.fused_moe.moe_load_balancer import \
     HostMoeTensorSharer
 
+pytestmark = pytest.mark.cpu_only
+
 
 class TestHostMoeTensorSharer(unittest.TestCase):
     """Tests for HostMoeTensorSharer functionality"""
diff --git a/tests/unittest/_torch/visual_gen/multi_gpu/test_attn2d_attention.py b/tests/unittest/_torch/visual_gen/multi_gpu/test_attn2d_attention.py
index dfb4061af984..dcdc4a023a45 100644
--- a/tests/unittest/_torch/visual_gen/multi_gpu/test_attn2d_attention.py
+++ b/tests/unittest/_torch/visual_gen/multi_gpu/test_attn2d_attention.py
@@ -831,6 +831,7 @@ def support_fused_qkv(cls):
         assert "head_dim" in str(e) or "num_heads" in str(e), f"Unexpected RuntimeError: {e}"
 
 
+@pytest.mark.cpu_only
 class TestAttn2DAttentionInitGuards:
     """Attention2DAttention.__init__ rejects invalid inner backends."""
 
diff --git a/tests/unittest/_torch/visual_gen/multi_gpu/test_visual_gen_mapping.py b/tests/unittest/_torch/visual_gen/multi_gpu/test_visual_gen_mapping.py
index c73762ae3c7a..df49725f1395 100644
--- a/tests/unittest/_torch/visual_gen/multi_gpu/test_visual_gen_mapping.py
+++ b/tests/unittest/_torch/visual_gen/multi_gpu/test_visual_gen_mapping.py
@@ -77,6 +77,7 @@ def _run_multi_gpu(world_size, test_fn):
 # =============================================================================
 
 
+@pytest.mark.cpu_only
 class TestConstruction:
     def test_single_gpu_defaults(self):
         vgm = VisualGenMapping(world_size=1, rank=0)
@@ -191,6 +192,7 @@ def test_ring_and_attn2d_raises(self):
             )
 
 
+@pytest.mark.cpu_only
 class TestFlattenCfgRanks:
     """flatten_cfg_ranks is pure layout arithmetic: one rank list per combined
     coordinate of every non-(cfg, ulysses) mesh dim, cfg outermost / ulysses
@@ -217,6 +219,7 @@ def test_cfg2_attn2d_2x2_u1(self):
         assert vgm.flatten_cfg_ranks() == [[0, 4], [1, 5], [2, 6], [3, 7]]
 
 
+@pytest.mark.cpu_only
 class TestSingleGPURanksAndGroups:
     def test_ranks_are_zero(self):
         vgm = VisualGenMapping(world_size=1, rank=0)
@@ -248,6 +251,7 @@ def test_attn2d_row_col_groups_none_without_attn2d(self):
         assert vgm.attn2d_col_group is None
 
 
+@pytest.mark.cpu_only
 class TestToLlmMapping:
     def test_single_gpu(self):
         vgm = VisualGenMapping(world_size=1, rank=0)
diff --git a/tests/unittest/_torch/visual_gen/multi_gpu/test_visual_gen_multinode.py b/tests/unittest/_torch/visual_gen/multi_gpu/test_visual_gen_multinode.py
index fddadb291d60..2944922b0b7f 100644
--- a/tests/unittest/_torch/visual_gen/multi_gpu/test_visual_gen_multinode.py
+++ b/tests/unittest/_torch/visual_gen/multi_gpu/test_visual_gen_multinode.py
@@ -15,6 +15,9 @@
 from tensorrt_llm.visual_gen.args import VisualGenArgs
 from tensorrt_llm.visual_gen.visual_gen import DiffusionRemoteClient, _detect_external_launch
 
+pytestmark = pytest.mark.cpu_only
+
+
 # =============================================================================
 # _detect_external_launch()
 # =============================================================================
diff --git a/tests/unittest/_torch/visual_gen/sparse_attention/test_skip_softmax.py b/tests/unittest/_torch/visual_gen/sparse_attention/test_skip_softmax.py
index 993ab9af0c5a..f4b757d49ba2 100644
--- a/tests/unittest/_torch/visual_gen/sparse_attention/test_skip_softmax.py
+++ b/tests/unittest/_torch/visual_gen/sparse_attention/test_skip_softmax.py
@@ -18,6 +18,8 @@
 from tensorrt_llm.visual_gen.args import AttentionConfig, VisualGenArgs
 from tensorrt_llm.visual_gen.sparse_attention import SkipSoftmaxAttentionConfig
 
+pytestmark = pytest.mark.cpu_only
+
 
 def _ckpt_sparse_attention_config(
     *,
diff --git a/tests/unittest/_torch/visual_gen/test_attention_integration.py b/tests/unittest/_torch/visual_gen/test_attention_integration.py
index a8babe53afb8..a372db9641f2 100644
--- a/tests/unittest/_torch/visual_gen/test_attention_integration.py
+++ b/tests/unittest/_torch/visual_gen/test_attention_integration.py
@@ -276,6 +276,7 @@ def generate_rope_embeddings(
 # ============================================================================
 
 
+@pytest.mark.cpu_only
 class TestSeparateQkvSequenceParallelGuard:
     def test_ring_with_separate_qkv_raises(self):
         vgm = VisualGenMapping(world_size=2, rank=0, ring_size=2)
diff --git a/tests/unittest/_torch/visual_gen/test_cache_dit.py b/tests/unittest/_torch/visual_gen/test_cache_dit.py
index 0301bb76b28a..e836cad6f995 100644
--- a/tests/unittest/_torch/visual_gen/test_cache_dit.py
+++ b/tests/unittest/_torch/visual_gen/test_cache_dit.py
@@ -241,6 +241,7 @@ def __init__(self, boundary_ratio: float | None, timestep_values: list[float]):
         self.transformer = torch.nn.Linear(1, 1)
 
 
+@pytest.mark.cpu_only
 class TestSplitWan22InferenceSteps:
     def test_counts_high_noise_at_default_boundary(self):
         pipeline = _MiniWan22Pipeline(
diff --git a/tests/unittest/_torch/visual_gen/test_flux_infer.py b/tests/unittest/_torch/visual_gen/test_flux_infer.py
index 87dd0f6a3d99..a50435a095ea 100644
--- a/tests/unittest/_torch/visual_gen/test_flux_infer.py
+++ b/tests/unittest/_torch/visual_gen/test_flux_infer.py
@@ -10,6 +10,8 @@
 
 from tensorrt_llm._torch.visual_gen.models.flux import Flux2Pipeline, FluxPipeline
 
+pytestmark = pytest.mark.cpu_only
+
 
 @pytest.mark.parametrize("pipeline_cls", [FluxPipeline, Flux2Pipeline])
 def test_infer_forwards_num_images_per_prompt(
diff --git a/tests/unittest/_torch/visual_gen/test_ltx2_pipeline.py b/tests/unittest/_torch/visual_gen/test_ltx2_pipeline.py
index 282fc81cf8c0..d0bb3fe43a69 100644
--- a/tests/unittest/_torch/visual_gen/test_ltx2_pipeline.py
+++ b/tests/unittest/_torch/visual_gen/test_ltx2_pipeline.py
@@ -478,6 +478,7 @@ def test_attention_backend_comparison(self, ltx2_bf16_checkpoint_exists):
 class TestLTX2BatchSupport:
     """Test batch support logic without loading the full pipeline."""
 
+    @pytest.mark.cpu_only
     def test_video_pixel_shape_batch_propagation(self):
         """VideoPixelShape(batch=N) propagates through VideoLatentShape."""
         from tensorrt_llm._torch.visual_gen.models.ltx2.ltx2_core.types import (
@@ -494,6 +495,7 @@ def test_video_pixel_shape_batch_propagation(self):
             torch_shape = video_shape.to_torch_shape()
             assert torch_shape[0] == batch_size
 
+    @pytest.mark.cpu_only
     def test_prompt_normalization(self):
         """forward() normalizes str prompt to List[str] and computes batch_size."""
         # Simulate the normalization logic from forward()
@@ -507,6 +509,7 @@ def test_prompt_normalization(self):
                 prompt = [prompt]
             assert len(prompt) == expected_batch
 
+    @pytest.mark.cpu_only
     def test_negative_prompt_expansion(self):
         """Negative prompt is expanded to match batch_size."""
         # Simulate the negative prompt expansion logic from forward()
@@ -548,6 +551,7 @@ def test_latent_shape_matches_batch(self):
 class TestTwoStageLoRAHelpers:
     """Test LoRA delta loading and application without checkpoints."""
 
+    @pytest.mark.cpu_only
     def test_bf16_weight_snapshot_gate_uses_cuda_free_memory(self, monkeypatch):
         from tensorrt_llm._torch.visual_gen.models.ltx2.pipeline_ltx2_two_stages import (
             _should_save_bf16_weights,
@@ -570,6 +574,7 @@ def _raise_mem_query_error(device=None):
         monkeypatch.setattr(torch.cuda, "is_available", lambda: False)
         assert not _should_save_bf16_weights()
 
+    @pytest.mark.cpu_only
     def test_bf16_weight_snapshot_saved_when_requested(self):
         from tensorrt_llm._torch.visual_gen.models.ltx2.pipeline_ltx2_two_stages import (
             _apply_lora_deltas,
@@ -808,6 +813,7 @@ def test_lora_strength_scaling(self):
             os.unlink(tmp_path)
 
 
+@pytest.mark.cpu_only
 class TestTwoStagePipelineVariantResolution:
     """Test that LTX2Pipeline.resolve_variant selects the correct class."""
 
@@ -875,6 +881,7 @@ def test_resolve_variant_requires_both_paths(self):
         assert result is LTX2Pipeline
 
 
+@pytest.mark.cpu_only
 class TestLTX2ForceOneStageEnv:
     """Test force-one-stage env-var behavior during LTX2 variant selection."""
 
@@ -1709,6 +1716,7 @@ def test_two_stage_with_quantization(self, ltx2_two_stage_assets_exist, quant_al
             torch.cuda.empty_cache()
 
 
+@pytest.mark.cpu_only
 class TestTwoStageCommonWarmupShapes:
     """Test LTX2TwoStagesPipeline.common_warmup_shapes property."""
 
diff --git a/tests/unittest/_torch/visual_gen/test_ltx2_transformer.py b/tests/unittest/_torch/visual_gen/test_ltx2_transformer.py
index 2cceb839f9a0..226623d4207b 100644
--- a/tests/unittest/_torch/visual_gen/test_ltx2_transformer.py
+++ b/tests/unittest/_torch/visual_gen/test_ltx2_transformer.py
@@ -585,6 +585,7 @@ class TestLTX2QuantExcludeModuleRemapping(unittest.TestCase):
     pre-quantized FP8 checkpoints are properly excluded from quantization.
     """
 
+    @pytest.mark.cpu_only
     def test_remap_qkv_fusion(self):
         """to_q/to_k/to_v should produce a qkv_proj entry."""
         from tensorrt_llm._torch.visual_gen.models.ltx2.transformer_ltx2 import LTXModel
@@ -597,6 +598,7 @@ def test_remap_qkv_fusion(self):
         remapped = LTXModel._remap_exclude_modules(exclude)
         self.assertIn("transformer_blocks.0.attn1.qkv_proj", remapped)
 
+    @pytest.mark.cpu_only
     def test_remap_ff(self):
         """ff.net.0.proj / ff.net.2 should produce up_proj / down_proj."""
         from tensorrt_llm._torch.visual_gen.models.ltx2.transformer_ltx2 import LTXModel
@@ -609,6 +611,7 @@ def test_remap_ff(self):
         self.assertIn("transformer_blocks.0.ff.up_proj", remapped)
         self.assertIn("transformer_blocks.0.ff.down_proj", remapped)
 
+    @pytest.mark.cpu_only
     def test_remap_audio_ff(self):
         """audio_ff.net.* should produce audio_ff.up_proj / down_proj."""
         from tensorrt_llm._torch.visual_gen.models.ltx2.transformer_ltx2 import LTXModel
@@ -621,6 +624,7 @@ def test_remap_audio_ff(self):
         self.assertIn("transformer_blocks.0.audio_ff.up_proj", remapped)
         self.assertIn("transformer_blocks.0.audio_ff.down_proj", remapped)
 
+    @pytest.mark.cpu_only
     def test_remap_preserves_originals(self):
         """Original entries (to_out.0, to_q for cross-attn) must survive."""
         from tensorrt_llm._torch.visual_gen.models.ltx2.transformer_ltx2 import LTXModel
diff --git a/tests/unittest/_torch/visual_gen/test_qwen_image_pipeline_config.py b/tests/unittest/_torch/visual_gen/test_qwen_image_pipeline_config.py
index db0f9a6db398..71ca19766751 100644
--- a/tests/unittest/_torch/visual_gen/test_qwen_image_pipeline_config.py
+++ b/tests/unittest/_torch/visual_gen/test_qwen_image_pipeline_config.py
@@ -52,6 +52,8 @@
     VisualGenArgs,
 )
 
+pytestmark = pytest.mark.cpu_only
+
 
 def _write_minimal_qwen_checkpoint(tmp_path: Path) -> Path:
     """Create the minimum diffusers layout needed by PipelineLoader config code."""
diff --git a/tests/unittest/_torch/visual_gen/test_teacache.py b/tests/unittest/_torch/visual_gen/test_teacache.py
index 2b6aa73519ad..5f8f3386105d 100644
--- a/tests/unittest/_torch/visual_gen/test_teacache.py
+++ b/tests/unittest/_torch/visual_gen/test_teacache.py
@@ -30,6 +30,8 @@
 from tensorrt_llm._torch.visual_gen.pipeline import BasePipeline
 from tensorrt_llm.visual_gen.args import TeaCacheConfig
 
+pytestmark = pytest.mark.cpu_only
+
 
 class _PipelineConfigShim:
     """Minimal pipeline_config shim that delegates all reads back to a DiffusionModelConfig.
diff --git a/tests/unittest/_torch/visual_gen/test_tensor_payload.py b/tests/unittest/_torch/visual_gen/test_tensor_payload.py
index 4ad482b886fd..6338f58989d0 100644
--- a/tests/unittest/_torch/visual_gen/test_tensor_payload.py
+++ b/tests/unittest/_torch/visual_gen/test_tensor_payload.py
@@ -22,6 +22,8 @@
 )
 from tensorrt_llm.visual_gen.output import VisualGenOutput
 
+pytestmark = pytest.mark.cpu_only
+
 
 def _safetensors_load(data: bytes) -> dict:
     from safetensors.torch import load as load_safetensors
diff --git a/tests/unittest/_torch/visual_gen/test_trtllm_serve_endpoints.py b/tests/unittest/_torch/visual_gen/test_trtllm_serve_endpoints.py
index 69d17ea523ef..eb343951e3a3 100644
--- a/tests/unittest/_torch/visual_gen/test_trtllm_serve_endpoints.py
+++ b/tests/unittest/_torch/visual_gen/test_trtllm_serve_endpoints.py
@@ -32,6 +32,9 @@
 from tensorrt_llm.serve.visual_gen_utils import VIDEO_STORE
 from tensorrt_llm.visual_gen.output import VisualGenMetrics, VisualGenOutput
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---------------------------------------------------------------------------
 # Helpers
 # ---------------------------------------------------------------------------
diff --git a/tests/unittest/_torch/visual_gen/test_utils.py b/tests/unittest/_torch/visual_gen/test_utils.py
index 384196e43e59..09705a5602c7 100644
--- a/tests/unittest/_torch/visual_gen/test_utils.py
+++ b/tests/unittest/_torch/visual_gen/test_utils.py
@@ -43,6 +43,7 @@ def _cleanup_mpi_env():
 # =============================================================================
 
 
+@pytest.mark.cpu_only
 class TestSequenceSharderInactive:
     def test_shard_gather_identity_when_size_one(self):
         s = SequenceSharder(size=1, rank=0, group=None)
@@ -63,6 +64,7 @@ def test_shard_rope_passthrough_when_inactive(self):
         assert s.shard_rope(rope, seq_len=4, seq_dim=1) is rope
 
 
+@pytest.mark.cpu_only
 class TestSequenceSharderShardSlices:
     """Active sharder: block slice math without ``gather``."""
 
@@ -86,6 +88,7 @@ def test_not_divisible_raises(self):
             s.shard(x, dim=1)
 
 
+@pytest.mark.cpu_only
 class TestSequenceSharderShardRope:
     def test_shard_rope_bsd_layout(self):
         s = SequenceSharder(size=2, rank=1, group=None)
@@ -112,6 +115,7 @@ def test_shard_rope_explicit_seq_dim_handles_square_layout(self):
         assert oc.shape == (2, 4, S)
 
 
+@pytest.mark.cpu_only
 class TestSequenceSharderFromVgm:
     def test_from_vgm_none(self):
         s = SequenceSharder.from_vgm(None)
@@ -177,6 +181,7 @@ def _run_dist(world_size: int, entry: Callable[[int, int, int], None]):
     mp.spawn(entry, args=(world_size, port), nprocs=world_size, join=True)
 
 
+@pytest.mark.cpu_only
 class TestSequenceSharderDistributed:
     def test_shard_gather_pad_unpad_four_ranks(self):
         _run_dist(4, _spawn_entry_combined)
diff --git a/tests/unittest/_torch/visual_gen/test_vanilla_key_padding_mask.py b/tests/unittest/_torch/visual_gen/test_vanilla_key_padding_mask.py
index 723126d97475..2cdf9919069b 100644
--- a/tests/unittest/_torch/visual_gen/test_vanilla_key_padding_mask.py
+++ b/tests/unittest/_torch/visual_gen/test_vanilla_key_padding_mask.py
@@ -13,6 +13,8 @@
 
 from tensorrt_llm._torch.visual_gen.attention_backend import VanillaAttention
 
+pytestmark = pytest.mark.cpu_only
+
 
 def test_padded_kv_with_mask_matches_unpadded():
     torch.manual_seed(42)
diff --git a/tests/unittest/_torch/visual_gen/test_visual_gen_args.py b/tests/unittest/_torch/visual_gen/test_visual_gen_args.py
index 13532eaa3458..49fffdfd9ff2 100644
--- a/tests/unittest/_torch/visual_gen/test_visual_gen_args.py
+++ b/tests/unittest/_torch/visual_gen/test_visual_gen_args.py
@@ -22,6 +22,8 @@
     VisualGenArgs,
 )
 
+pytestmark = pytest.mark.cpu_only
+
 
 class TestVisualGenArgsStrictValidation:
     """extra='forbid' rejects unknown fields at every nesting level."""
diff --git a/tests/unittest/_torch/visual_gen/test_visual_gen_params.py b/tests/unittest/_torch/visual_gen/test_visual_gen_params.py
index cfa43240458b..9d1f95d83d71 100644
--- a/tests/unittest/_torch/visual_gen/test_visual_gen_params.py
+++ b/tests/unittest/_torch/visual_gen/test_visual_gen_params.py
@@ -18,6 +18,8 @@
 
 import pytest
 
+pytestmark = pytest.mark.cpu_only
+
 
 def _wan_mock(is_wan22_14b=False, is_wan22_5b=False, name_or_path="", num_heads=12):
     """Create a mock with attributes needed by WanPipeline/WanI2V properties."""
diff --git a/tests/unittest/_torch/visual_gen/test_visual_gen_utils.py b/tests/unittest/_torch/visual_gen/test_visual_gen_utils.py
index d6bf720b12f2..ea2a96e6f12a 100644
--- a/tests/unittest/_torch/visual_gen/test_visual_gen_utils.py
+++ b/tests/unittest/_torch/visual_gen/test_visual_gen_utils.py
@@ -29,6 +29,8 @@
 )
 from tensorrt_llm.visual_gen import VisualGenParams
 
+pytestmark = pytest.mark.cpu_only
+
 
 class _StubExtraParamSpec:
     def __init__(self, default: Any = None) -> None:
diff --git a/tests/unittest/_torch/weight_sharing/test_gms_source_identity_gate.py b/tests/unittest/_torch/weight_sharing/test_gms_source_identity_gate.py
index 99f21b4bd9b9..e67260daa657 100644
--- a/tests/unittest/_torch/weight_sharing/test_gms_source_identity_gate.py
+++ b/tests/unittest/_torch/weight_sharing/test_gms_source_identity_gate.py
@@ -29,6 +29,8 @@
 from tensorrt_llm._torch.weight_sharing import SourceIdentityMismatchError
 from tensorrt_llm.llmapi.llm_args import LoadFormat
 
+pytestmark = pytest.mark.cpu_only
+
 
 class _FakeCheckpointLoader:
     """Minimal checkpoint-loader stub exposing `checkpoint_format`."""
diff --git a/tests/unittest/_torch/weight_sharing/test_mx_source_identity_gate.py b/tests/unittest/_torch/weight_sharing/test_mx_source_identity_gate.py
index e2658ca86d21..a5cf9c87c717 100644
--- a/tests/unittest/_torch/weight_sharing/test_mx_source_identity_gate.py
+++ b/tests/unittest/_torch/weight_sharing/test_mx_source_identity_gate.py
@@ -21,6 +21,7 @@
 
 from types import SimpleNamespace
 
+import pytest
 from _source_identity_fakes import FakeMapping
 from _source_identity_fakes import make_identity as _identity
 
@@ -29,6 +30,8 @@
     _build_mx_source_metadata,
 )
 
+pytestmark = pytest.mark.cpu_only
+
 
 def _new_loader(local_identity):
     """Construct a loader while bypassing the heavy base initializer."""
diff --git a/tests/unittest/_torch/weight_sharing/test_source_identity.py b/tests/unittest/_torch/weight_sharing/test_source_identity.py
index 6d574f351a61..da3671d7397a 100644
--- a/tests/unittest/_torch/weight_sharing/test_source_identity.py
+++ b/tests/unittest/_torch/weight_sharing/test_source_identity.py
@@ -41,6 +41,8 @@
     check_weight_sharing_compatibility,
 )
 
+pytestmark = pytest.mark.cpu_only
+
 
 def test_identical_configs_match():
     a = identity_from(FakeModelConfig())
diff --git a/tests/unittest/api_stability/test_llm_api.py b/tests/unittest/api_stability/test_llm_api.py
index 259eab8b62cc..e6b59903e80a 100644
--- a/tests/unittest/api_stability/test_llm_api.py
+++ b/tests/unittest/api_stability/test_llm_api.py
@@ -14,6 +14,8 @@
 from tensorrt_llm.sampling_params import (BatchedLogitsProcessor,
                                           LogitsProcessor, SamplingParams)
 
+pytestmark = pytest.mark.cpu_only
+
 
 class TestSamplingParams(ApiStabilityTestHarness):
     TEST_CLASS = SamplingParams
diff --git a/tests/unittest/dynamo/test_imports.py b/tests/unittest/dynamo/test_imports.py
index 8e29ff34127a..888da5ddb1fd 100644
--- a/tests/unittest/dynamo/test_imports.py
+++ b/tests/unittest/dynamo/test_imports.py
@@ -34,6 +34,9 @@
 
 import pytest
 
+pytestmark = pytest.mark.cpu_only
+
+
 # Each entry is (module_path, symbol_name).  The test ID shows both.
 DYNAMO_IMPORTS = [
     # -- top-level --
diff --git a/tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_manager_v2.py b/tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_manager_v2.py
index 48c10a006fc0..7668e9fb8a5e 100755
--- a/tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_manager_v2.py
+++ b/tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_manager_v2.py
@@ -28,6 +28,8 @@
 from statistics import median
 from typing import TYPE_CHECKING, Any, Iterator, NamedTuple, cast, get_type_hints
 
+import pytest
+
 if not TYPE_CHECKING and find_spec("kv_cache_manager_v2") is not None:
     from kv_cache_manager_v2 import (
         DEFAULT_BEAM_INDEX,
@@ -4302,6 +4304,7 @@ def test_shrink_touched_pool(self) -> None:
             allocator.release(s)
 
 
+@pytest.mark.cpu_only
 class TestBlockKeyHashing(unittest.TestCase):
     """Verify Hasher.update produces bit-identical digests to the per-token reference (no GPU needed)."""
 
diff --git a/tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_salting.py b/tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_salting.py
index 4c2dc9615686..17c404d2139e 100644
--- a/tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_salting.py
+++ b/tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_salting.py
@@ -19,6 +19,11 @@
 from importlib.util import find_spec
 from typing import TYPE_CHECKING, cast
 
+import pytest
+
+pytestmark = pytest.mark.cpu_only
+
+
 if not TYPE_CHECKING and find_spec("kv_cache_manager_v2") is not None:
     from kv_cache_manager_v2 import TokenId, sequence_to_blockchain_keys
     from kv_cache_manager_v2._block_radix_tree import Block, BlockRadixTree, ReuseScope
diff --git a/tests/unittest/scaffolding/test_bench.py b/tests/unittest/scaffolding/test_bench.py
index 95dc29862fd5..8d51b7764ef6 100644
--- a/tests/unittest/scaffolding/test_bench.py
+++ b/tests/unittest/scaffolding/test_bench.py
@@ -17,6 +17,8 @@
     ConcurrentStrategy, ConstantRateStrategy, LoadGenerationStrategy,
     PoissonRateStrategy, SynchronousStrategy, ThroughputStrategy)
 
+pytestmark = pytest.mark.cpu_only
+
 OUTPUT_STR = "Yes."
 
 
diff --git a/tests/unittest/scaffolding/test_mcp_worker.py b/tests/unittest/scaffolding/test_mcp_worker.py
index 7faabff62769..4716782c3118 100644
--- a/tests/unittest/scaffolding/test_mcp_worker.py
+++ b/tests/unittest/scaffolding/test_mcp_worker.py
@@ -26,6 +26,9 @@
 )
 from tensorrt_llm.scaffolding.worker import MCPWorker, Worker
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ============================================================
 # MCP Server Definition (based on websearch.py)
 # ============================================================
diff --git a/tests/unittest/scaffolding/test_parallel_process.py b/tests/unittest/scaffolding/test_parallel_process.py
index e277b9d97acd..c5088056aff5 100644
--- a/tests/unittest/scaffolding/test_parallel_process.py
+++ b/tests/unittest/scaffolding/test_parallel_process.py
@@ -4,9 +4,13 @@
 from enum import Enum
 from typing import List
 
+import pytest
+
 from tensorrt_llm.scaffolding import (Controller, ParallelProcess,
                                       ScaffoldingLlm, Task, TaskStatus, Worker)
 
+pytestmark = pytest.mark.cpu_only
+
 
 class DummyTask(Task):
 
diff --git a/tests/unittest/scaffolding/test_task_collection.py b/tests/unittest/scaffolding/test_task_collection.py
index aa5fd483e807..6e35833a9565 100644
--- a/tests/unittest/scaffolding/test_task_collection.py
+++ b/tests/unittest/scaffolding/test_task_collection.py
@@ -2,6 +2,8 @@
 from enum import Enum
 from typing import List
 
+import pytest
+
 from tensorrt_llm.scaffolding import (AssistantMessage, ChatTask, Controller,
                                       DropKVCacheTask, DropKVCacheWorkerTag,
                                       ParallelProcess, ScaffoldingLlm, Task,
@@ -9,6 +11,8 @@
                                       Worker, drop_kv_cache_scope,
                                       sub_request_node, with_task_collection)
 
+pytestmark = pytest.mark.cpu_only
+
 
 class DummyTask(Task):
 
diff --git a/tests/unittest/usage/test_collectors.py b/tests/unittest/usage/test_collectors.py
index a8b8d0038991..4053e631b40a 100644
--- a/tests/unittest/usage/test_collectors.py
+++ b/tests/unittest/usage/test_collectors.py
@@ -17,8 +17,13 @@
 import json
 from unittest.mock import MagicMock, patch
 
+import pytest
+
 from tensorrt_llm.usage import schema, usage_lib
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---------------------------------------------------------------------------
 # System info tests
 # ---------------------------------------------------------------------------
diff --git a/tests/unittest/usage/test_config.py b/tests/unittest/usage/test_config.py
index 9083aba69825..f3714ae728f9 100644
--- a/tests/unittest/usage/test_config.py
+++ b/tests/unittest/usage/test_config.py
@@ -16,6 +16,8 @@
 
 import pytest
 
+pytestmark = pytest.mark.cpu_only
+
 
 class TestTelemetryConfigLocation:
     """Verify TelemetryConfig and UsageContext live in tensorrt_llm.usage.config."""
diff --git a/tests/unittest/usage/test_llmapi_config_capture.py b/tests/unittest/usage/test_llmapi_config_capture.py
index 37bf95636560..dae4be48512b 100644
--- a/tests/unittest/usage/test_llmapi_config_capture.py
+++ b/tests/unittest/usage/test_llmapi_config_capture.py
@@ -18,6 +18,8 @@
 from pathlib import Path
 from typing import Any, Literal, Optional, Union
 
+import pytest
+
 from tensorrt_llm.llmapi.llm_args import (
     CudaGraphConfig,
     Field,
@@ -29,6 +31,8 @@
 from tensorrt_llm.usage import usage_lib
 from tensorrt_llm.usage.llmapi_config import collect_llm_api_config_payloads
 
+pytestmark = pytest.mark.cpu_only
+
 
 class _NestedConfig(StrictBaseModel):
     marked: int = Field(default=7, telemetry={"kind": "value"})
diff --git a/tests/unittest/usage/test_llmapi_config_telemetry_docs.py b/tests/unittest/usage/test_llmapi_config_telemetry_docs.py
index 1d0dbf89918f..059fad4224ef 100644
--- a/tests/unittest/usage/test_llmapi_config_telemetry_docs.py
+++ b/tests/unittest/usage/test_llmapi_config_telemetry_docs.py
@@ -21,6 +21,10 @@
 from pathlib import Path
 from types import ModuleType
 
+import pytest
+
+pytestmark = pytest.mark.cpu_only
+
 
 def _repo_root() -> Path:
     return Path(__file__).resolve().parents[3]
diff --git a/tests/unittest/usage/test_opt_out.py b/tests/unittest/usage/test_opt_out.py
index f88fe8126c5a..67b15b8f9c46 100644
--- a/tests/unittest/usage/test_opt_out.py
+++ b/tests/unittest/usage/test_opt_out.py
@@ -21,6 +21,9 @@
 
 from tensorrt_llm.usage import usage_lib
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---------------------------------------------------------------------------
 # Opt-out tests
 # ---------------------------------------------------------------------------
diff --git a/tests/unittest/usage/test_reporter.py b/tests/unittest/usage/test_reporter.py
index 130b08a54b68..53b74bf3dfce 100644
--- a/tests/unittest/usage/test_reporter.py
+++ b/tests/unittest/usage/test_reporter.py
@@ -20,10 +20,14 @@
 from types import SimpleNamespace
 from unittest.mock import MagicMock, patch
 
+import pytest
 from pydantic import BaseModel, Field
 
 from tensorrt_llm.usage import usage_lib
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---------------------------------------------------------------------------
 # Console notification tests
 # ---------------------------------------------------------------------------
diff --git a/tests/unittest/usage/test_schema.py b/tests/unittest/usage/test_schema.py
index 1603b3550762..f22a9ccfe3db 100644
--- a/tests/unittest/usage/test_schema.py
+++ b/tests/unittest/usage/test_schema.py
@@ -21,6 +21,9 @@
 
 from tensorrt_llm.usage import schema, schemas
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---------------------------------------------------------------------------
 # Features JSON payload structure tests
 # ---------------------------------------------------------------------------
diff --git a/tests/unittest/usage/test_transport.py b/tests/unittest/usage/test_transport.py
index 878d9a18ae88..bdbce27b39f7 100644
--- a/tests/unittest/usage/test_transport.py
+++ b/tests/unittest/usage/test_transport.py
@@ -27,6 +27,7 @@
 # ---------------------------------------------------------------------------
 
 
+@pytest.mark.cpu_only
 class TestSendToGxt:
     def test_send_fail_silent(self):
         """_send_to_gxt never raises on network error."""
@@ -105,6 +106,7 @@ def test_send_to_gxt_catches_url_error(self, monkeypatch):
 # ---------------------------------------------------------------------------
 
 
+@pytest.mark.cpu_only
 class TestHttpsHandler:
     def test_opener_has_https_handler(self):
         """Opener includes HTTPSHandler for HTTPS endpoints."""
@@ -133,6 +135,7 @@ def mock_build_opener(*handlers):
 # ---------------------------------------------------------------------------
 
 
+@pytest.mark.cpu_only
 class TestMalformedServerUrl:
     """Verify _send_to_gxt handles malformed server URLs without crashing."""
 
diff --git a/tests/unittest/visual_gen/test_output.py b/tests/unittest/visual_gen/test_output.py
index a4e0540d624e..73dbb32ce9dd 100644
--- a/tests/unittest/visual_gen/test_output.py
+++ b/tests/unittest/visual_gen/test_output.py
@@ -18,6 +18,9 @@
 )
 from tensorrt_llm.visual_gen import VisualGenMetrics, VisualGenOutput
 
+pytestmark = pytest.mark.cpu_only
+
+
 # ---------------------------------------------------------------------------
 # VisualGenOutput shape and re-exports
 # ---------------------------------------------------------------------------