Point to tekit_backend (#6616)

* Point to tekit_backend

* Update version

Author: krishung5

build.py

@@ -1319,18 +1319,50 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach
     # Add dependencies needed for tensorrtllm backend
     if "tensorrtllm" in backends:
         be = "tensorrtllm"
-        url = "https://raw.githubusercontent.com/triton-inference-server/tensorrtllm_backend/{}/tools/gen_trtllm_dockerfile.py".format(
-            backends[be]
-        )
-
-        response = requests.get(url)
-        spec = importlib.util.spec_from_loader(
-            "trtllm_buildscript", loader=None, origin=url
+        df += """
+WORKDIR /workspace
+RUN apt-get update && apt-get install -y --no-install-recommends python3-pip
+
+# Remove previous TRT installation
+RUN apt-get remove --purge -y tensorrt* libnvinfer*
+RUN pip uninstall -y tensorrt
+# Install new version of TRT using the script from TRT-LLM
+RUN apt-get update && apt-get install -y --no-install-recommends python-is-python3
+RUN git clone --single-branch --depth=1 -b {} https://{}:{}@gitlab-master.nvidia.com/ftp/tekit_backend.git tensorrtllm_backend
+RUN cd tensorrtllm_backend && git submodule update --init --recursive
+ENV TRT_VER=9.2.0.4
+ENV CUDA_VER=12.3
+ENV CUDNN_VER=8.9.6.50-1+cuda12.2
+ENV NCCL_VER=2.19.3-1+cuda12.3
+ENV CUBLAS_VER=12.3.2.9-1
+RUN cp tensorrtllm_backend/tensorrt_llm/docker/common/install_tensorrt.sh /tmp/
+RUN rm -fr tensorrtllm_backend
+    """.format(
+            backends[be],
+            os.environ["REMOVE_ME_TRTLLM_USERNAME"],
+            os.environ["REMOVE_ME_TRTLLM_TOKEN"],
         )
-        trtllm_buildscript = importlib.util.module_from_spec(spec)
-        exec(response.content, trtllm_buildscript.__dict__)
-        df += trtllm_buildscript.create_postbuild(backends[be])
 
+        df += """
+RUN bash /tmp/install_tensorrt.sh --CUDA_VER=$CUDA_VER --CUDNN_VER=$CUDNN_VER --NCCL_VER=$NCCL_VER --CUBLAS_VER=$CUBLAS_VER && rm /tmp/install_tensorrt.sh
+ENV TRT_ROOT=/usr/local/tensorrt
+# Remove TRT contents that are not needed in runtime
+RUN ARCH="$(uname -i)" && \
+    rm -fr ${TRT_ROOT}/bin ${TRT_ROOT}/targets/${ARCH}-linux-gnu/bin ${TRT_ROOT}/data && \
+    rm -fr  ${TRT_ROOT}/doc ${TRT_ROOT}/onnx_graphsurgeon ${TRT_ROOT}/python && \
+    rm -fr ${TRT_ROOT}/samples  ${TRT_ROOT}/targets/${ARCH}-linux-gnu/samples
+
+# Install required packages for TRT-LLM models
+RUN python3 -m pip install --upgrade pip && \
+        pip3 install transformers
+
+# Uninstall unused nvidia packages
+RUN if pip freeze | grep -q "nvidia.*"; then \
+        pip freeze | grep "nvidia.*" | xargs pip uninstall -y; \
+    fi
+RUN pip cache purge
+ENV LD_LIBRARY_PATH=/usr/local/tensorrt/lib/:/opt/tritonserver/backends/tensorrtllm:$LD_LIBRARY_PATH
+"""
     if "vllm" in backends:
         # [DLIS-5606] Build Conda environment for vLLM backend
         # Remove Pip install once vLLM backend moves to Conda environment.
@@ -1790,6 +1822,12 @@ def core_build(
 
 def tensorrtllm_prebuild(cmake_script):
     # Export the TRT_ROOT environment variable
+    cmake_script.cmd("export TRT_VER=9.2.0.4")
+    cmake_script.cmd("export CUDA_VER=12.3")
+    cmake_script.cmd("export CUDNN_VER=8.9.6.50-1+cuda12.2")
+    cmake_script.cmd("export NCCL_VER=2.19.3-1+cuda12.3")
+    cmake_script.cmd("export CUBLAS_VER=12.3.2.9-1")
+
     cmake_script.cmd("export TRT_ROOT=/usr/local/tensorrt")
     cmake_script.cmd("export ARCH=$(uname -m)")
 
@@ -1820,10 +1858,18 @@ def backend_build(
     cmake_script.comment()
     cmake_script.mkdir(build_dir)
     cmake_script.cwd(build_dir)
-    cmake_script.gitclone(backend_repo(be), tag, be, github_organization)
 
     if be == "tensorrtllm":
+        cmake_script.cmd(
+            "git clone --single-branch --depth=1 -b {} https://{}:{}@gitlab-master.nvidia.com/ftp/tekit_backend.git tensorrtllm".format(
+                tag,
+                os.environ["REMOVE_ME_TRTLLM_USERNAME"],
+                os.environ["REMOVE_ME_TRTLLM_TOKEN"],
+            )
+        )
         tensorrtllm_prebuild(cmake_script)
+    else:
+        cmake_script.gitclone(backend_repo(be), tag, be, github_organization)
 
     cmake_script.mkdir(repo_build_dir)
     cmake_script.cwd(repo_build_dir)