[TRTLLM-13696][test] Part2.4: Migrate CPU only tests - others

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:

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) ---------------
@@ -195,19 +194,14 @@ 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
   - unittest/_torch/visual_gen/test_attention_cute_dsl_vsa.py
   - 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

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

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

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]

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"

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

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

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."""

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):

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 = {

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
 # ---------------------------------------------------------------------------

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
 
 

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
 # ---------------------------------------------------------------------------

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)
 

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
 # ---------------------------------------------------------------------------

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
 # --------------------------------------------------------------------------- #

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 ----
 
 

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``.

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"
 
 

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"""

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."""