[None][chore] Lock onnx version <1.20.0 and remove WAR for TRT 10.13 (#9006)

Signed-off-by: Yiqing Yan <[email protected]>
Signed-off-by: Yanchao Lu <[email protected]>
Co-authored-by: Yanchao Lu <[email protected]>

Author: yiqingy0

docs/source/installation/linux.md

@@ -59,3 +59,15 @@ There are some known limitations when you pip install pre-built TensorRT LLM whe
     when OMPI was not configured --with-slurm and we weren't able
     to discover a SLURM installation in the usual places.
     ```
+
+2. Prevent `pip` from replacing existing PyTorch installation
+
+   On certain systems, particularly Ubuntu 22.04, users installing TensorRT LLM would find that their existing, CUDA 13.0 compatible PyTorch installation (e.g., `torch==2.9.0+cu130`) was being uninstalled by `pip`. It was then replaced by a CUDA 12.8 version (`torch==2.9.0`), causing the TensorRT LLM installation to be unusable and leading to runtime errors.
+
+   The solution is to create a `pip` constraints file, locking `torch` to the currently installed version. Here is an example of how this can be done manually:
+
+   ```bash
+   CURRENT_TORCH_VERSION=$(python3 -c "import torch; print(torch.__version__)")
+   echo "torch==$CURRENT_TORCH_VERSION" > /tmp/torch-constraint.txt
+   pip3 install --upgrade pip setuptools && pip3 install tensorrt_llm -c /tmp/torch-constraint.txt
+   ```

requirements.txt

@@ -8,7 +8,7 @@ diffusers>=0.27.0
 lark
 mpi4py
 numpy<2
-onnx>=1.18.0
+onnx>=1.18.0,<1.20.0
 onnx_graphsurgeon>=0.5.2
 openai
 polygraphy
@@ -19,8 +19,6 @@ pandas
 h5py==3.12.1
 StrEnum
 sentencepiece>=0.1.99
-# WAR for tensorrt depending on the archived nvidia-cuda-runtime-cu13 package
-nvidia-cuda-runtime-cu13==0.0.0a0
 tensorrt~=10.13.0
 # https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-25-10.html#rel-25-10 uses 2.9.0a0.
 torch>=2.9.0a0,<=2.9.0

tests/unittest/test_pip_install.py

@@ -105,6 +105,51 @@ def download_wheel(args):
                           shell=True)
 
 
+def install_tensorrt_llm():
+    """
+    Installs the tensorrt_llm wheel, dynamically creating a torch constraint
+    if torch is already installed to prevent it from being replaced.
+    """
+    print("##########  Install tensorrt_llm package  ##########")
+
+    install_command = "pip3 install tensorrt_llm-*.whl"
+
+    # Always check for an existing torch installation, regardless of OS.
+    try:
+        print("Checking for existing torch installation...")
+        torch_version_result = subprocess.run(
+            ['python3', '-c', 'import torch; print(torch.__version__)'],
+            capture_output=True,
+            text=True,
+            check=True)
+        torch_version = torch_version_result.stdout.strip()
+
+        if torch_version:
+            print(f"Found installed torch version: {torch_version}")
+            constraint_filename = "torch-constraint.txt"
+            with open(constraint_filename, "w") as f:
+                f.write(f"torch=={torch_version}\n")
+            print(
+                f"Created {constraint_filename} to constrain torch to version {torch_version}."
+            )
+
+            # Modify install command to use the constraint
+            install_command += f" -c {constraint_filename}"
+        else:
+            # This case is unlikely if the subprocess call succeeds
+            print(
+                "Could not determine installed torch version. Installing without constraint."
+            )
+
+    except (subprocess.CalledProcessError, FileNotFoundError):
+        # This handles cases where python3 fails or 'import torch' raises an error.
+        print("Torch is not installed. Proceeding without constraint.")
+
+    # Execute the final installation command
+    print(f"Executing command: {install_command}")
+    subprocess.check_call(install_command, shell=True)
+
+
 def test_pip_install():
     parser = argparse.ArgumentParser(description="Check Pip Install")
     parser.add_argument("--wheel_path",
@@ -125,8 +170,8 @@ def test_pip_install():
                           shell=True)
 
     download_wheel(args)
-    print("##########  Install tensorrt_llm package  ##########")
-    subprocess.check_call("pip3 install tensorrt_llm-*.whl", shell=True)
+    install_tensorrt_llm()
+
     print("##########  Test import tensorrt_llm  ##########")
     subprocess.check_call(
         'python3 -c "import tensorrt_llm; print(tensorrt_llm.__version__)"',