[6034518] Fix remote safety benchmark latency

CHANGELOG.rst

@@ -133,6 +133,7 @@ Changelog
   tensors; use resident export for those models.
 - Fix shared ONNX export metadata and Diffusers attention policy: every ``NVFP4QuantExporter`` post-process now upgrades the default-domain opset to at least 23, all FP8 custom-op exports re-run ONNX shape/type inference after setting output metadata, and quantized SDPA derives FP8 MHA enablement from the live Q/K/V quantizers instead of honoring a caller-set ``_disable_fp8_mha`` attribute.
 - Fix ONNX FP16 conversion failing to preserve public output types when type inference changes a graph output declaration before output casts are inserted.
+- Fix ONNX remote AutoQDQ safety benchmarks returning infinite latency by running the generated engine with ``trtexec_safe`` on the configured target.
 - Fix ``examples/hf_ptq/hf_ptq.py`` discarding a completed PTQ run (no checkpoint exported) when the optional post-quantization sanity-check ``generate()`` call raised, for example because ``device_map="auto"`` placed part of the model on CPU. That failure is now caught and only skips the sanity check; export proceeds regardless.
 - Fix ``examples/megatron_bridge/export_quantized_megatron_to_hf.py`` storing the MoE router at Megatron's ``moe_router_dtype``, which is a routing *compute* dtype, not a storage one. The router now exports at the export ``dtype`` like every other unquantized weight, matching what ``hf_ptq.py`` and the released NVFP4 checkpoints contain; pass ``moe_router_dtype`` to ``export_mcore_gpt_to_hf`` explicitly if you want the old fp32 storage.
 - Fix unified Megatron export writing a second, unreferenced copy of the vocab embedding when a model with MTP layers is exported with pipeline parallelism. The duplicate was never loaded but inflated the checkpoint by the size of the embedding (about 1 GB for Qwen3.6-35B-A3B); re-export to reclaim the space.

docs/source/guides/9_autotune.rst

@@ -251,11 +251,17 @@ To use remote autotuning during Q/DQ placement optimization, run with ``trtexec`
 **Requirements:**
 
 * TensorRT 10.15 or later
-* Valid remote autotuning configuration
+* Valid ``ssh://`` remote autotuning configuration without a password
+* Non-interactive SSH key authentication from the host to the target
+* ``trtexec_safe`` on the target, alongside the configured ``remote_exec_path``
 * ``--use_trtexec`` must be set (benchmarking uses ``trtexec`` instead of the TensorRT Python API)
 * ``--safe --skipInference`` must be enabled via ``--trtexec_benchmark_args``
 
-Replace ``<remote autotuning config>`` with an actual remote autotuning configuration string (see ``trtexec --help`` for more details). Other TensorRT benchmark options (e.g. ``--timing_cache``, ``--warmup_runs``, ``--timing_runs``, ``--plugin_libraries``) are also available; run ``--help`` for details.
+Replace ``<remote autotuning config>`` with an actual remote autotuning configuration string (see ``trtexec --help`` for more details). ModelOpt uses the configuration to build the engine with remote autotuning, copies the generated engine to an internal temporary path on the target, and runs ``trtexec_safe`` there to measure GPU compute time. ModelOpt attempts to remove the temporary engine after each benchmark.
+
+Other TensorRT benchmark options (e.g. ``--timing_cache``, ``--warmup_runs``, ``--timing_runs``, ``--plugin_libraries``) are also available; run ``--help`` for details.
+
+``--plugin_libraries`` applies to the host-side engine build only. ModelOpt does not transfer or load custom plugin libraries on the remote target, so this workflow does not support plugin-dependent remote safety engines.
 
 Low-Level API Usage
 ===================

examples/onnx_ptq/autotune/README.md

@@ -247,12 +247,17 @@ python3 -m modelopt.onnx.quantization.autotune \
 **Requirements:**
 
 - TensorRT 10.15 or later
-- Valid remote autotuning configuration
+- Valid `ssh://` remote autotuning configuration without a password
+- Non-interactive SSH key authentication from the host to the target
+- `trtexec_safe` on the target, alongside the configured `remote_exec_path`
 - `--use_trtexec` must be set (benchmarking uses `trtexec` instead of the TensorRT Python API)
 - `--safe --skipInference` must be enabled via `--trtexec_benchmark_args`
 
-Replace `<remote autotuning config>` with an actual remote autotuning configuration string (see `trtexec --help` for more details).
- Other TensorRT benchmark options (e.g. `--timing_cache`, `--warmup_runs`, `--timing_runs`, `--plugin_libraries`) are also available; run `--help` for details.
+Replace `<remote autotuning config>` with an actual remote autotuning configuration string (see `trtexec --help` for more details). ModelOpt uses the configuration to build the engine with remote autotuning, copies the generated engine to an internal temporary path on the target, and runs `trtexec_safe` there to measure GPU compute time. ModelOpt attempts to remove the temporary engine after each benchmark.
+
+Other TensorRT benchmark options (e.g. `--timing_cache`, `--warmup_runs`, `--timing_runs`, `--plugin_libraries`) are also available; run `--help` for details.
+
+`--plugin_libraries` applies to the host-side engine build only. ModelOpt does not transfer or load custom plugin libraries on the remote target, so this workflow does not support plugin-dependent remote safety engines.
 
 ## Programmatic API Usage