NV embedding layer loading

[z52527]

Problem

Some PT2/AOTI models need a one-time native initialization at model load — before the package is loaded — that cannot live inside model.pt2. A concrete example: a model whose custom ops resolve embedding tables from a process-global registry at execute time, where the weights are stored in sidecar files next to the package (not inside model.pt2). Loading the .pt2 does not load them, so the first inference request fails:

... no binding registered for layer_id=1
-> Inductor model instance execution failure ... model_container_runner.cpp:150

The Triton PyTorch backend currently has no entry point to run such load-time initialization, so today it can only be done with fragile workarounds (e.g. an LD_PRELOAD shim that hijacks an unrelated CUDA call to piggyback on its timing).

What this PR does

Adds a generic, optional model-init hook to the PT2 path. When a model's config.pbtxt sets the MODEL_INIT_LIBRARY parameter, the backend dlopen()s that shared library in LoadModel() — right before the AOTIModelPackageLoader is created (CUDA is already initialized there) — and calls a fixed C entry point. Design choices:

• Backend stays generic — it does not link against the library and knows nothing about what it does. It only dlopens the .so, calls triton_pytorch_model_init(model_dir, device_index), and releases it (triton_pytorch_model_release + dlclose) on unload. All model-specific logic lives in the user-provided plug-in, built and shipped downstream.
• No new dependency — the only added link is ${CMAKE_DL_LIBS} (CMake's built-in libdl variable, for dlopen/dlsym/dlclose). The backend links no vendor libraries, so the stock nvcr.io/nvidia/pytorch:26.05-py3 build is unaffected.
• No-op by default — a model without the parameter is completely unaffected; a dlopen/dlsym failure fails model load (fail-closed).
• Scoped to the model lifetime — the handle is held in ModelState and freed in the destructor.
• Fixes a pre-existing parameters bug — ParseParameters() was guarded by if (!ModelConfig().Find("parameters", ...)), so PT2 parameters were parsed only when the section was absent; any model that set a parameters block had them all (INFERENCE_MODE, …) silently ignored. Dropping the ! makes them take effect.

The library must export:

// Called once at model load, before the package is loaded.
void* triton_pytorch_model_init(const char* model_dir, int device_index);
// Optional; called once on unload with the handle returned above.
void  triton_pytorch_model_release(void* state);

Verified end-to-end on the official nvcr.io/nvidia/tritonserver:26.05-py3 image with an AOTI model (parameter parsed → hook loaded → model READY → inference correct); also compiles against main.

[whoisj]

Overall, this is good. Thank you for the contribution. I've left a couple of comments but am willing to accept this change as-is.

[mc-nv]

CMAKE_ prefix is a reserved keyword, afraid it's usage may create some confusion. See examples below.
http://cmake.org/cmake/help/latest/manual/cmake-variables.7.html

Also change description explains following, which means those libraries are vital to preserve the functionality, isn't?

An HSTU AOTI package (model.pt2) calls nve_ops::embedding_lookup(keys, layer_id),
which looks the embedding table up by layer_id in a process-global
NVELayerRegistry. The embedding weights do not live inside model.pt2 —
they are separate files next to it (<model_dir>/metadata.json +
<model_dir>/weights/*.nve). So loading the .pt2 does not load them, and the
first inference request fails

Could you please share the missed libraries origin ?

[whoisj]

@z52527 please follow up with @mc-nv's requests. This is blocking merge.

[z52527]

Sorry for the confusion. PR description was stale; it still described the earlier "link NVE into the backend" version. I've updated it.
In the current revision the backend links no NVE or vendor libraries. There are no missing libraries to source. The only CMake change is target_link_libraries(... ${CMAKE_DL_LIBS}). CMAKE_DL_LIBS is a built-in CMake variable for the platform's dynamic-loading library (libdl), needed for the dlopen/dlsym/dlclose the hook uses.

[whoisj]

CI Pipeline ID: 55028851

[whoisj]

I've pushed a change to make the pre-commit hook happy. @z52527, please pull your branch to receive my contribution before working on it to avoid conflicts; or duplicate my work in your own commit. Thanks.