[None][fix] AutoDeploy: Unwaive llmc standalone tests

examples/auto_deploy/llmc/create_standalone_package.py

@@ -158,10 +158,17 @@
     "test_flashinfer_mamba_cached_op.py",
     # Require TRT-LLM custom ops (dsv3_router_gemm_op, noaux_tc_op, etc.)
     "test_deepseek_custom.py",
+    "test_glm4_moe_modeling.py",
     "test_glm4_moe_lite_modeling.py",
+    "test_glm_moe_dsa_modeling.py",
+    # Full-model tests hit standalone-incompatible HF cache behavior.
+    "test_granite_moe_hybrid_modeling.py",
+    # Imports triton_kernels, which is not a standalone dependency.
+    "test_mxfp4_moe_layout.py",
     # Require TRT-LLM distributed ops (trtllm_dist_all_gather)
     "test_gather_logits_before_lm_head.py",
-    # Multimodal types are None in standalone (MultimodalInput guard)
+    # Multimodal processors depend on TensorRT-LLM multimodal request types.
+    "test_gemma4_modeling.py",
     "test_qwen3_5_moe.py",
     # Hardware-specific (requires H100+ shared memory)
     "test_triton_mla_op.py",
@@ -377,8 +384,16 @@ def _create_test_conftest(tests_dir: str) -> None:
         # limitations under the License.
 
         \"\"\"Conftest for standalone auto_deploy tests.\"\"\"
-        import sys
+        import importlib.util
         import os
+        import sys
+
+        _trtllm_spec = importlib.util.find_spec("tensorrt_llm")
+        if _trtllm_spec is not None:
+            raise RuntimeError(
+                "Standalone llmc tests must not be able to import tensorrt_llm; "
+                f"found {getattr(_trtllm_spec, 'origin', None)!r}"
+            )
 
         # Add _utils_test to the Python path so test files can import from it
         sys.path.insert(0, os.path.join(os.path.dirname(__file__), "_utils_test"))

tensorrt_llm/_torch/auto_deploy/custom_ops/normalization/rms_norm.py

@@ -21,10 +21,13 @@
 import torch.nn.functional as F
 from einops import rearrange
 
-import tensorrt_llm.quantization.utils.fp4_utils as fp4_utils
-
 from ..quantization.quant import TRTLLM_NVFP4_SCALING_VECTOR_SIZE
 
+try:
+    import tensorrt_llm.quantization.utils.fp4_utils as fp4_utils
+except (ModuleNotFoundError, ImportError):
+    fp4_utils = None
+
 try:
     from tensorrt_llm._torch.flashinfer_utils import get_env_enable_pdl
 except (ModuleNotFoundError, ImportError):
@@ -41,8 +44,22 @@ def get_env_enable_pdl() -> bool:
 from .triton_rms_norm import rms_norm
 
 
+def _pad_up(x, y: int):
+    return ((x + y - 1) // y) * y
+
+
 def _get_nvfp4_fake_shapes(x: torch.Tensor) -> tuple[tuple[int, ...], int]:
-    output_shape, sf_size = fp4_utils.get_fp4_shape(x.shape, TRTLLM_NVFP4_SCALING_VECTOR_SIZE)
+    if fp4_utils is not None:
+        output_shape, sf_size = fp4_utils.get_fp4_shape(x.shape, TRTLLM_NVFP4_SCALING_VECTOR_SIZE)
+        return tuple(output_shape), sf_size
+
+    input_shape = tuple(x.shape)
+    m = 1
+    for dim in input_shape[:-1]:
+        m *= dim
+    output_shape = list(input_shape)
+    output_shape[-1] //= 2
+    sf_size = _pad_up(m, 128) * _pad_up(input_shape[-1] // TRTLLM_NVFP4_SCALING_VECTOR_SIZE, 4)
     return tuple(output_shape), sf_size
 
 

tensorrt_llm/_torch/auto_deploy/models/custom/modeling_minimax_m2.py

@@ -36,8 +36,7 @@
 from transformers.modeling_utils import PreTrainedModel
 from transformers.utils import ModelOutput
 
-from tensorrt_llm._torch.utils import ActivationType
-
+from ..._compat import ActivationType
 from ..hf import AutoModelForCausalLMFactory
 from .rotary_utils import RotaryEmbeddingBase, build_rope_cos_sin_cache
 

tensorrt_llm/_torch/auto_deploy/models/quant_config_reader.py

@@ -25,13 +25,29 @@
 from abc import ABC, abstractmethod
 from typing import Any, Callable, Dict, Optional, Tuple, Type
 
-from tensorrt_llm.quantization.modelopt_config import (
-    is_modelopt_quant_config,
-    read_modelopt_quant_config,
-)
-
+from .._compat import TRTLLM_AVAILABLE
 from ..utils.logger import ad_logger
 
+# Importing ``tensorrt_llm.quantization.modelopt_config`` triggers
+# ``tensorrt_llm.quantization.__init__`` which pulls in ``tensorrt_llm.mapping``
+# and thus ``tensorrt_llm._torch.device_mesh``; none of these exist in the
+# standalone llmc package. Gate the import so this module loads cleanly when
+# TRT-LLM is absent; the ModelOPT reader path itself is unreachable in that case.
+if TRTLLM_AVAILABLE:
+    from tensorrt_llm.quantization.modelopt_config import (
+        is_modelopt_quant_config,
+        read_modelopt_quant_config,
+    )
+else:
+
+    def is_modelopt_quant_config(raw: Any) -> bool:
+        return False
+
+    def read_modelopt_quant_config(raw: Any) -> Dict[str, Any]:
+        raise NotImplementedError(
+            "ModelOPT quant config parsing requires TensorRT-LLM; not available in standalone mode."
+        )
+
 
 class QuantConfigReader(ABC):
     """Base class for reading and parsing quantization config."""

tensorrt_llm/_torch/auto_deploy/transform/library/fuse_rope_mla.py

@@ -35,13 +35,14 @@
 import torch
 from torch.fx import GraphModule, Node
 
-from ...custom_ops.mla.trtllm_mla import _TRTLLM_MLA_ROPE_INFO_KEY
 from ...models.factory import ModelFactory
 from ...shim.interface import CachedSequenceInterface
 from ...utils.logger import ad_logger
 from ...utils.node_utils import is_op
 from ..interface import BaseTransform, SharedConfig, TransformInfo, TransformRegistry
 
+_TRTLLM_MLA_ROPE_INFO_KEY = "_trtllm_mla_rope_info"
+
 # At post_load_fusion, only the backend-agnostic torch_rope_* IR ops are
 # present (optimize_rope has not yet replaced them with flashinfer_rope).
 _ROPE_OP_TARGETS = []

tensorrt_llm/_torch/auto_deploy/transform/library/sharding_ir.py

@@ -111,6 +111,13 @@ def _shard_scale_and_hook(
     }
 )
 
+
+def _auto_deploy_ops(*names: str) -> Tuple[OpOverloadPacket, ...]:
+    return tuple(
+        op for name in names if (op := getattr(torch.ops.auto_deploy, name, None)) is not None
+    )
+
+
 # =============================================================================
 # ShardableNode abstract base class
 # =============================================================================
@@ -524,10 +531,7 @@ def apply(self, gm: GraphModule, dc: DistConfig, max_num_tokens: int = 0) -> int
         return 1 if count > 0 else 0
 
 
-@ShardableNode.register(
-    torch.ops.auto_deploy.torch_rmsnorm_gated,
-    torch.ops.auto_deploy.triton_rmsnorm_gated,
-)
+@ShardableNode.register(*_auto_deploy_ops("torch_rmsnorm_gated", "triton_rmsnorm_gated"))
 class NormShardableNode(ShardableNode):
     """Gated RMSNorm op: shard weight parameter."""
 

tensorrt_llm/_torch/auto_deploy/utils/_graph.py

@@ -357,7 +357,27 @@ def lint(gm: GraphModule) -> None:
     gm.graph.lint()
 
 
+def _restore_topological_order(gm: GraphModule) -> None:
+    """Move nodes after their direct inputs if graph rewrites left them out of order."""
+    while True:
+        node_order = {node: idx for idx, node in enumerate(gm.graph.nodes)}
+        for node in list(gm.graph.nodes):
+            input_nodes = [
+                input_node for input_node in node.all_input_nodes if input_node in node_order
+            ]
+            if not input_nodes:
+                continue
+            latest_input = max(input_nodes, key=node_order.__getitem__)
+            if node_order[node] < node_order[latest_input]:
+                latest_input.append(node)
+                break
+        else:
+            return
+
+
 def _canonicalize_single_gm(gm: GraphModule) -> None:
+    _restore_topological_order(gm)
+
     # clean up graph (needs to be done repeatedly until no more dead code)
     eliminate_dead_code(gm, is_impure_node=_is_impure_node)
 

tensorrt_llm/_torch/auto_deploy/utils/node_utils.py

@@ -46,6 +46,10 @@
 OperatorLike = Union[OpOrOverload, Callable]
 
 
+def _auto_deploy_op(name: str) -> Optional[OpOverloadPacket]:
+    return getattr(torch.ops.auto_deploy, name, None)
+
+
 class LayerType(Enum):
     """Enum for layer type."""
 
@@ -670,13 +674,17 @@ def is_any_moe_op(node: Node) -> bool:
     return is_op(
         node,
         ops=[
-            torch.ops.auto_deploy.torch_moe,
-            torch.ops.auto_deploy.torch_quant_fp8_moe,
-            torch.ops.auto_deploy.torch_quant_nvfp4_moe,
-            torch.ops.auto_deploy.torch_quant_finegrained_fp8_moe,
-            torch.ops.auto_deploy.triton_mxfp4_moe,
-            torch.ops.auto_deploy.torch_moe_fused,
-            torch.ops.auto_deploy.torch_moe_dense_mlp,
+            op
+            for op in [
+                _auto_deploy_op("torch_moe"),
+                _auto_deploy_op("torch_quant_fp8_moe"),
+                _auto_deploy_op("torch_quant_nvfp4_moe"),
+                _auto_deploy_op("torch_quant_finegrained_fp8_moe"),
+                _auto_deploy_op("triton_mxfp4_moe"),
+                _auto_deploy_op("torch_moe_fused"),
+                _auto_deploy_op("torch_moe_dense_mlp"),
+            ]
+            if op is not None
         ],
     )
 

tests/integration/test_lists/waives.txt

@@ -359,8 +359,6 @@ unittest/auto_deploy/singlegpu/models/test_qwen3_5_moe.py::test_vision_attention
 unittest/auto_deploy/singlegpu/models/test_qwen3_5_moe.py::test_vision_block_matches_reference SKIP (https://nvbugs/6189450)
 unittest/auto_deploy/singlegpu/models/test_qwen3_5_moe.py::test_vlm_wrapper_delta_is_request_scoped_no_cross_call_leakage SKIP (https://nvbugs/6189450)
 unittest/auto_deploy/singlegpu/smoke/test_ad_build_small_single.py::test_build_ad[deepseek-ai/DeepSeek-V3-llm_extra_args10] SKIP (https://nvbugs/5888827)
-unittest/auto_deploy/standalone SKIP (https://nvbugs/6160629)
-unittest/auto_deploy/standalone/test_standalone_package.py::TestStandalonePackage::test_run_unit_tests SKIP (https://nvbugs/6160629)
 unittest/disaggregated/test_agent_multi_backends.py::test_run_with_different_env[1] SKIP (https://nvbugs/5979673)
 unittest/executor/test_rpc.py::TestRpcCorrectness::test_incremental_task_async SKIP (https://nvbugs/5741476)
 unittest/executor/test_rpc_proxy.py SKIP (https://nvbugs/5605741)

tests/unittest/auto_deploy/standalone/test_standalone_package.py

@@ -174,27 +174,65 @@ def test_run_unit_tests(self, standalone_package):
         if not os.path.isdir(tests_dir):
             pytest.skip("No tests directory in standalone package")
 
-        # Pass through the host env but override PYTHONPATH to use standalone tests.
-        # The venv's pip install already put llmc in the venv's site-packages,
-        # so `import llmc` resolves there (not to the host TRT-LLM).
+        # Keep runtime/compiler environment, but remove host Python import paths.
         standalone_env = {
-            **os.environ,
-            "PYTHONPATH": tests_dir + os.pathsep + os.path.join(tests_dir, "_utils_test"),
-            # Override PATH to prefer venv's python/pytest
-            "PATH": os.path.join(standalone_package["venv_dir"], "bin")
-            + os.pathsep
-            + os.environ.get("PATH", ""),
-            # FlashInfer JIT-compiles kernels and caches .so files under
-            # FLASHINFER_WORKSPACE_BASE (default: $HOME). Ninja records absolute
-            # source paths from the flashinfer package. Since this venv is
-            # ephemeral, a subsequent run would find stale cache entries pointing
-            # at the old (deleted) venv paths, causing all JIT builds to fail.
-            # Redirect the cache into the venv so it's discarded with it.
-            "FLASHINFER_WORKSPACE_BASE": standalone_package["venv_dir"],
+            key: value
+            for key, value in os.environ.items()
+            if key not in {"PYTHONHOME", "PYTHONPATH", "PYTHONUSERBASE"}
         }
+        standalone_env.update(
+            {
+                # Override PATH to prefer venv's python/pytest.
+                "PATH": os.path.join(standalone_package["venv_dir"], "bin")
+                + os.pathsep
+                + os.environ.get("PATH", ""),
+                # Keep subprocess workers from adding cwd/user paths that can
+                # expose the TensorRT-LLM checkout to standalone tests.
+                "PYTHONNOUSERSITE": "1",
+                "PYTHONSAFEPATH": "1",
+                # FlashInfer JIT-compiles kernels and caches .so files under
+                # FLASHINFER_WORKSPACE_BASE (default: $HOME). Ninja records absolute
+                # source paths from the flashinfer package. Since this venv is
+                # ephemeral, a subsequent run would find stale cache entries pointing
+                # at the old (deleted) venv paths, causing all JIT builds to fail.
+                # Redirect the cache into the venv so it's discarded with it.
+                "FLASHINFER_WORKSPACE_BASE": standalone_package["venv_dir"],
+            }
+        )
+
+        isolation_probe = subprocess.run(
+            [
+                python,
+                "-I",
+                "-c",
+                textwrap.dedent(
+                    """
+                    import importlib.util
+                    import sys
+
+                    spec = importlib.util.find_spec("tensorrt_llm")
+                    if spec is not None:
+                        raise SystemExit(
+                            "tensorrt_llm is importable in standalone test env: "
+                            "origin=%r, sys.path=%r" % (getattr(spec, "origin", None), sys.path)
+                        )
+                    """
+                ),
+            ],
+            capture_output=True,
+            text=True,
+            timeout=30,
+            cwd=pkg_dir,
+            env=standalone_env,
+        )
+        assert isolation_probe.returncode == 0, (
+            f"Standalone env leaked TensorRT-LLM\n"
+            f"stdout:\n{isolation_probe.stdout}\nstderr:\n{isolation_probe.stderr}"
+        )
 
         cmd = [
             python,
+            "-I",
             "-m",
             "pytest",
             os.path.join(tests_dir, "singlegpu"),