[Bug][Installation]: `TRTLLM_USE_PRECOMPILED=1` fails sanity check with Python 3.12 bindings

System Info

Branch: release/1.1.0rc3
Docker Image: nvcr.io/nvidia/tensorrt-llm/devel:1.1.0rc3
Python Version: 3.12

How you are installing TensorRT-LLM

TRTLLM_USE_PRECOMPILED=1 pip wheel . --no-deps --wheel-dir ./build

Description

The TRTLLM_USE_PRECOMPILED=1 flag fails to skip C++ compilation despite documentation stating it should download and extract precompiled libraries. The installation fails at the sanity check stage, even though the precompiled bindings are successfully extracted.

Root Cause

The sanity_check() function checks for:tensorrt_llm/bindings or tensorrt_llm/bindings.pyi.
However, the precompiled wheel extracts: bindings.cpython-312-x86_64-linux-gnu.so (compiled shared object)
The sanity check logic doesn't recognize the platform-specific .so file pattern, causing it to incorrectly fail even when the bindings are present.

Steps to Reproduce

1. Start devel container: `make -C docker ngc-devel_run LOCAL_USER=1 DOCKER_PULL=1`
2. Run: `TRTLLM_USE_PRECOMPILED=1 pip install -e . -v`
3. Observe sanity check failure
4. Verify bindings exist: `find . -name "bindings*.so" -type f`

Expected Behavior

Installation should succeed without requiring python3 ./scripts/build_wheel.py when TRTLLM_USE_PRECOMPILED=1 is set.

Suggested Fix

Update sanity_check() to also check for platform-specific bindings patterns like bindings.cpython-*.so or bindings*.so.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the installation documentation and examples for answers to frequently asked questions.