[TRTLLM-13696][test] Part2.1: Migrate CPU only tests - runtime

jenkins/L0_Test.groovy

@@ -1393,7 +1393,7 @@ def getPytestBaseCommandLine(
     if (stageName.contains("-Ray-")) {
         testCmdLine += ["--run-ray"]
     }
-    def unittestMarkExpr = (stageName.startsWith("CPU-")) ? "cpu_only and not disabled" : "not cpu_only"
+    def unittestMarkExpr = (stageName.startsWith("CPU-")) ? "cpu_only" : "not cpu_only"
     testCmdLine += ["--unittest-markexpr='${unittestMarkExpr}'"]
     if (ENABLE_UPLOAD_TEST_RESULTS) {
         testCmdLine += ["-o console_output_style=progress-even-when-capture-no"]
@@ -5075,7 +5075,7 @@ def launchTestJobs(pipeline, testFilter)
     // may break the mapping functionality.
 
     x86TestConfigs = [
-        "CPU-Generic-x86-1": ["cpu", "l0_cpu_x86", 1, 1],
+        "CPU-Generic-x86-1": ["cpu", "l0_cpu", 1, 1],
         "DGX_H100-4_GPUs-CPP-1": ["dgx-h100-x4", "l0_dgx_h100", 1, 1, 4],
         "A10-PyTorch-1": ["a10", "l0_a10", 1, 3],
         "A10-PyTorch-2": ["a10", "l0_a10", 2, 3],
@@ -5257,7 +5257,7 @@ def launchTestJobs(pipeline, testFilter)
 
     // SBSA machines from the Blossom machine pool
     SBSATestConfigs = [
-        "CPU-Generic-arm-1": ["cpu", "l0_cpu_arm", 1, 1],
+        "CPU-Generic-arm-1": ["cpu", "l0_cpu", 1, 1],
         "GH200-PyTorch-Post-Merge-1": ["gh200", "l0_gh200", 1, 1],
         // DGX Spark is also named as GB10 Grace Blackwell Superchip.
         "GB10-PyTorch-1": ["gb10x", "l0_gb10", 1, 1],

tensorrt_llm/llmapi/llm_args.py

@@ -4368,8 +4368,7 @@ class BaseLlmArgs(StrictBaseModel):
         "(default), it is read from the HF config.json ('dtype', or the "
         "deprecated 'torch_dtype'); for composite/VLM configs it falls "
         "back to the nested text_config.dtype. Defaults to bfloat16 if "
-        "none is found, and is overridden to float16 on GPUs with compute "
-        "capability < 8.0 (pre-Ampere).",
+        "none is found.",
         telemetry=TelemetryField.categorical("auto", "float16", "bfloat16",
                                              "float32"))
 
@@ -4679,16 +4678,6 @@ def from_yaml(cls, yaml_path: Union[str, Path]):
             raise ValueError("Configuration file root must be a mapping.")
         return cls(**config_dict)
 
-    @field_validator("dtype")
-    @classmethod
-    def validate_dtype(cls, v, info):
-        if torch.cuda.get_device_properties(0).major < 8:
-            if v == 'auto':
-                v = 'float16'
-            if v == 'bfloat16':
-                raise RuntimeError("Pre SM 80 GPUs do not support bfloat16")
-        return v
-
     @field_validator("gpus_per_node", mode='before')
     @classmethod
     def validate_gpus_per_node(cls, v, info):

tests/README.md

@@ -123,11 +123,13 @@ In file `jenkins/L0_Test.groovy`, the variables `x86TestConfigs`, `SBSATestConfi
 
 Currently the yml files are manually maintained, which requires developer to update them when new test cases are added.
 
-### How to choose GPU type
+### How to choose whether a GPU is needed and which GPU type
 
-The CI resource of each GPU type is different. Usually you should choose the cheapest GPU that fulfills test requirements. In most cases, an integration test case should only run on one GPU type, unless it's very important or has different behaviours on different GPUs.
+While TensorRT LLM is GPU-centric, the specific feature covered by a unit test may not need to interact with a GPU on its own. For unit tests that do not intrinsically require a GPU, prefer the CPU-only stages to avoid reserving unnecessary GPU resources. The CPU-only option applies only to unit tests.
 
-The priority is A10 > A30 > L40s > A100 > H100 > B200.
+When a test requires a GPU, choose the cheapest GPU that fulfills its requirements. In most cases, an integration test case should run on only one GPU type, unless it is especially important or behaves differently on different GPUs.
+
+The priority is A10 > A30 > L40S > A100 > H100 > B200.
 
 ## 2. Add an integration test
 
@@ -141,6 +143,8 @@ Once a new integration test case is added, the yml files must be updated to cont
 
 A unit test are used to test a standalone feature or building block, and only runs partial workflow.
 
+Mark a unit test that does not require a GPU to run with `@pytest.mark.cpu_only` or file-level `pytestmark = pytest.mark.cpu_only`. Only these tests run in the CPU-only stages. You can define CPU and GPU cases in the same test file.
+
 For legacy and case management reason, the CI doesn't run unit tests directly. It uses a bridge to map multiple unit test cases into one integration test case, and manages these bridged cases.
 The bridge is implemented in `integration/defs/test_unittests.py` and `pytest_generate_tests` function in `tests/integration/defs/conftest.py`.
 
@@ -161,9 +165,9 @@ pytest unittest/an_existing_file.py -m "part0 and gpu2" # run some cases in a fi
 ```
 
 2. Check existing bridge cases and make sure your cases are not covered by an existing one.
-For example, you may want to add `pytest unittest/an_existing_file.py -k "some_keyword or another_keyword"`, but there is already `pytest unittest/an_existing_file.py -k "not thrid_keyword"` which covers your filter.
+For example, you may want to add `pytest unittest/an_existing_file.py -k "some_keyword or another_keyword"`, but there is already `pytest unittest/an_existing_file.py -k "not third_keyword"` which covers your filter.
 
-3. Choose a suitable GPU and add a line of your cases. For example, adding `unittest/an_existing_file.py -k "some_keyword or another_keyword"` to `tests/integration/test_lists/test-db/l0_a10.yml`.
+3. Choose a suitable CI stage and add a line to the corresponding test list config. For example, adding `unittest/an_existing_file.py -k "some_keyword or another_keyword"` to `tests/integration/test_lists/test-db/l0_a10.yml`.
 
 ## 4. Run a CI stage locally
 

tests/integration/defs/test_unittests.py

@@ -131,6 +131,14 @@ def test_unittests_v2(llm_root, llm_venv, case: str, output_dir, request):
 
     import shlex
     arg_list = shlex.split(case)
+    if unittest_markexpr:
+        if "-m" in arg_list:
+            markexpr_index = arg_list.index("-m") + 1
+            case_markexpr = arg_list[markexpr_index]
+            arg_list[markexpr_index] = (
+                f"({case_markexpr}) and ({unittest_markexpr})")
+        else:
+            arg_list += ["-m", unittest_markexpr]
     case_fn = re.sub(r'[/\s"\']+', '-', case)
     if len(case_fn) > 80:
         case_fn = case_fn[:80]
@@ -171,9 +179,6 @@ def test_unittests_v2(llm_root, llm_venv, case: str, output_dir, request):
     if run_ray:
         command += ["--run-ray"]
 
-    if unittest_markexpr:
-        command += ["-m", unittest_markexpr]
-
     s3_secret_key = None
     s3_upload_path = request.config.getoption("--s3-upload-path", default=None)
     if s3_upload_path:

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

@@ -34,8 +34,6 @@ 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_async_transfer_manager.py
-  - unittest/_torch/executor/test_scheduler_serializable_output.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
@@ -44,21 +42,13 @@ l0_a10:
   - 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/dwdp/test_dwdp_manager.py
-  - unittest/_torch/modules/dwdp/test_dwdp_mapping.py
-  - unittest/_torch/modules/dwdp/test_dwdp_peer_ranges.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).
   - unittest/_torch/models/checkpoints
   - unittest/_torch/models/test_qwen3_next_moe_quant.py
   - unittest/_torch/weight_sharing
-  - unittest/inputs/test_chat_template_dispatch.py
-  - unittest/inputs/test_content_format.py
-  - unittest/inputs/test_url_validation.py
-  - unittest/inputs/test_multimodal.py
   - unittest/inputs/test_multimodal_input_processor.py
-  - unittest/inputs/test_video_decode.py
   - unittest/others/test_cache_transceiver_precheck_config.py
   - unittest/others/test_cache_transceiver_precheck_run.py
   - unittest/others/test_convert_utils.py
@@ -145,26 +135,14 @@ l0_a10:
   - unittest/_torch/visual_gen/test_qwen_image_infer.py
   - unittest/_torch/visual_gen/test_qwen_image_pipeline.py
   # llmapi
-  - unittest/llmapi/test_llm_utils.py
-  - unittest/llmapi/test_gc_utils.py
-  - unittest/llmapi/test_reasoning_parser.py
-  - unittest/llmapi/test_serialization.py
-  - unittest/llmapi/test_utils.py
   - unittest/llmapi/test_rlhf_utils.py
   - unittest/llmapi/test_llm_args.py
-  - unittest/llmapi/test_kv_cache_dtype_override.py
   - unittest/llmapi/test_additional_model_outputs.py -m "gpu1"
-  - unittest/llmapi/test_request_priority.py
   # executor
-  - unittest/executor/test_ipc.py
-  - unittest/executor/test_fatal_error_health_check.py
   - unittest/executor/test_postprocessor_hook.py
   - unittest/executor/test_proxy_postproc_terminate.py
   - unittest/executor/test_proxy_fast_death.py
   # trtllm-serve CPU-only
-  - unittest/llmapi/apps/test_chat_utils.py
-  - unittest/llmapi/apps/test_tool_parsers.py
-  - unittest/llmapi/apps/test_harmony_channel_validation.py
   - unittest/llmapi/apps/test_encode_batcher.py
   - unittest/llmapi/test_embedding_arch_routing.py
   # usage telemetry

tests/integration/test_lists/test-db/l0_a100.yml

@@ -56,7 +56,6 @@ l0_a100:
   tests:
   - unittest/llmapi/test_llm.py -m "part0"
   - unittest/llmapi/test_llm.py -m "not part0" TIMEOUT (90)
-  - unittest/llmapi/test_executor.py
 - condition:
     ranges:
       system_gpu_count:

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/executor/test_stats_serializer.py
   - 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
@@ -269,7 +268,6 @@ l0_b200:
       stage: post_merge
       backend: pytorch
   tests:
-  - unittest/llmapi/test_llm_quant.py # 3.5 mins on B200
   - unittest/disaggregated/test_openai_server_info.py
   - examples/visual_gen/test_visual_gen.py::test_cosmos3_t2i_4step_example TIMEOUT (30)
   - examples/visual_gen/test_visual_gen.py::test_cosmos3_i2v_4step_example TIMEOUT (45)

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

@@ -0,0 +1,46 @@
+version: 0.0.1
+l0_cpu:
+- condition:
+    ranges:
+      system_gpu_count:
+        gte: 0
+        lte: 0
+    wildcards:
+      linux_distribution_name: ubuntu*
+    terms:
+      stage: pre_merge
+      backend: generic
+      orchestrator: mpi
+  tests:
+  - unittest/_torch/distributed
+  - unittest/_torch/executor
+  - unittest/_torch/lora
+  - unittest/_torch/modules
+  - unittest/_torch/ray_orchestrator/single_gpu/test_cache_transceiver_comm.py
+  - unittest/_torch/speculative/hw_agnostic
+  - unittest/executor/test_base_worker.py ISOLATION
+  - unittest/executor/test_fatal_error_health_check.py
+  - unittest/executor/test_ipc.py
+  - unittest/executor/test_rpc.py
+  - unittest/executor/test_multi_frontend_routing.py
+  - unittest/executor/test_event_loop_error_broadcast.py
+  - unittest/executor/test_stats_serializer.py
+  - unittest/inputs
+  - unittest/llmapi/apps/test_chat_utils.py
+  - unittest/llmapi/apps/test_harmony_channel_validation.py
+  - unittest/llmapi/apps/test_tool_parsers.py
+  - unittest/llmapi/test_bench_async.py
+  - unittest/llmapi/test_additional_model_outputs.py -m "gpu1"
+  - unittest/llmapi/test_executor.py
+  - unittest/llmapi/test_gc_utils.py
+  - unittest/llmapi/test_kv_cache_dtype_override.py
+  - unittest/llmapi/test_llm_args.py
+  - unittest/llmapi/test_llm_quant.py
+  - unittest/llmapi/test_llm_telemetry.py
+  - unittest/llmapi/test_llm_utils.py
+  - unittest/llmapi/test_mpi_session.py ISOLATION
+  - unittest/llmapi/test_reasoning_parser.py
+  - unittest/llmapi/test_request_priority.py
+  - unittest/llmapi/test_serialization.py
+  - unittest/llmapi/test_utils.py
+  - unittest/others/test_http_utils_fail_fast.py

tests/integration/test_lists/test-db/l0_dgx_h100.yml

@@ -18,7 +18,6 @@ l0_dgx_h100:
   - unittest/llmapi/test_llm_multi_gpu_pytorch.py -m "gpu2"
   - unittest/llmapi/test_additional_model_outputs.py -m "gpu2"
   - unittest/_torch/multi_gpu -m "not post_merge" TIMEOUT (90)
-  - unittest/_torch/distributed
   - unittest/_torch/modeling/test_modeling_pixtral.py::test_tensor_parallelism
   - kv_cache/test_final_single_token_context_cuda_graph.py::test_final_token_reuse_cuda_graph_tp2[v1]
   - kv_cache/test_final_single_token_context_cuda_graph.py::test_final_token_reuse_cuda_graph_tp2[v2]
@@ -63,7 +62,6 @@ l0_dgx_h100:
   - unittest/llmapi/apps/test_disagg_serving_perf_metrics.py
   - disaggregated/test_disaggregated.py::test_disaggregated_cancel_large_context_requests[DeepSeek-V3-Lite-bf16]
   # llmapi
-  - unittest/llmapi/test_mpi_session.py::test_llmapi_launch_multiple_tasks
   - accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_gen_only_spec_dec
 - condition:
     ranges:

tests/integration/test_lists/test-db/l0_gh200.yml

@@ -15,7 +15,6 @@ l0_gh200:
       backend: pytorch
   tests:
   - unittest/bindings
-  - unittest/llmapi/test_llm_quant.py
   - llmapi/test_llm_examples.py::test_llmapi_quickstart_atexit
   - examples/visual_gen/test_visual_gen.py::test_visual_gen_quickstart
   - examples/visual_gen/test_visual_gen.py::test_visual_gen_api_walkthrough

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

@@ -268,7 +268,6 @@ l0_h100:
     - disaggregated/test_disaggregated_single_gpu.py::test_disaggregated_logprobs[False-TinyLlama-1.1B-Chat-v1.0]
     - disaggregated/test_disaggregated_single_gpu.py::test_disaggregated_logprobs[True-TinyLlama-1.1B-Chat-v1.0]
     - unittest/_torch/executor/test_overlap_scheduler.py
-    - unittest/_torch/ray_orchestrator/single_gpu/test_cache_transceiver_comm.py
     - unittest/_torch/ray_orchestrator/single_gpu/test_llm_sleep.py
     - unittest/_torch/ray_orchestrator/single_gpu/test_llm_update_weights.py -m "part0"
     - unittest/_torch/ray_orchestrator/single_gpu/test_llm_update_weights.py -m "part1"
@@ -310,7 +309,6 @@ l0_h100:
       stage: pre_merge
       backend: pytorch
   tests:
-  - unittest/llmapi/test_llm_quant.py # 5.5 mins on H100
   - examples/visual_gen/test_visual_gen.py::test_visual_gen_quickstart
   - examples/visual_gen/test_visual_gen.py::test_visual_gen_api_walkthrough
 - condition:
@@ -364,7 +362,6 @@ l0_h100:
   - unittest/bindings # 8 mins on H100
   - unittest/kv_cache_manager_v2_tests # 4 min
   # ------------- KV Cache Iteration Stats ---------------
-  - unittest/executor/test_stats_serializer.py
   - 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

tests/unittest/_torch/distributed/test_cp_broadcast.py

@@ -16,6 +16,8 @@
 from tensorrt_llm._torch.distributed import MPIDist
 from tensorrt_llm.mapping import Mapping
 
+pytestmark = pytest.mark.cpu_only
+
 
 def get_mpi_info():
     """Get MPI rank and world size, returns (0, 1) if MPI is not available."""

tests/unittest/_torch/distributed/test_safe_mpi_comm.py

@@ -34,6 +34,8 @@
 from tensorrt_llm._torch.distributed import communicator
 from tensorrt_llm.bindings import BuildInfo
 
+pytestmark = pytest.mark.cpu_only
+
 
 def get_mpi_info():
     """Get MPI rank and world size, returns (0, 1) if MPI is not available."""

tests/unittest/_torch/executor/test_adp_router.py

@@ -26,6 +26,8 @@
 from tensorrt_llm.conversation_params import ConversationParams
 from tensorrt_llm.scheduling_params import SchedulingParams
 
+pytestmark = pytest.mark.cpu_only
+
 
 class _MockRequest(MagicMock):
     """Mock executor Request whose ``num_input_tokens`` mirrors the real binding

tests/unittest/_torch/executor/test_async_transfer_manager.py

@@ -15,10 +15,14 @@
 
 from unittest.mock import MagicMock
 
+import pytest
+
 from tensorrt_llm._torch.pyexecutor.py_executor import AsyncTransferManager
 from tensorrt_llm._torch.pyexecutor.resource_manager import ResourceManagerType
 from tensorrt_llm.bindings import LlmRequestState
 
+pytestmark = pytest.mark.cpu_only
+
 
 def create_mock_request(request_id: int):
     """Create a mock LlmRequest with the given request ID."""

tests/unittest/_torch/executor/test_executor_request_queue.py

@@ -18,6 +18,8 @@
 from tensorrt_llm._torch.pyexecutor.executor_request_queue import (
     SHUTDOWN_REQUEST_ID, ExecutorRequestQueue, RequestQueueItem)
 
+pytestmark = pytest.mark.cpu_only
+
 
 @pytest.fixture
 def mock_dist():