Deploy NeMo Models by Exporting TensorRT-LLM#

This section shows how to use scripts and APIs to export a NeMo LLM (.nemo) or quantized LLM (.qnemo) to TensorRT-LLM and deploy it with the NVIDIA Triton Inference Server.

Quick Example#

  1. Follow the steps in the Deploy NeMo LLM main page to generate a NeMo 2.0 Llama checkpoint.

  2. In a terminal, go to the folder where the hf_llama31_8B_nemo2.nemo file is located. Pull down and run the Docker container image using the command shown below. Change the :vr tag to the version of the container you want to use:

    docker pull nvcr.io/nvidia/nemo:vr

docker run --gpus all -it --rm --shm-size=4g -p 8000:8000 \
    -v ${PWD}/hf_llama31_8B_nemo2.nemo:/opt/checkpoints/hf_llama31_8B_nemo2.nemo \
    -w /opt/NeMo \
    --name nemo-fw \
    nvcr.io/nvidia/nemo:vr

  3. Install TensorRT-LLM by executing the following command inside the container:

    cd /opt/Export-Deploy
uv sync --link-mode symlink --locked --extra trtllm $(cat /opt/uv_args.txt)

  4. Run the following deployment script to verify that everything is working correctly. The script exports the Llama NeMo checkpoint to TensorRT-LLM and subsequently serves it on the Triton server:

    python /opt/Export-Deploy/scripts/deploy/nlp/deploy_triton.py \
    --nemo_checkpoint /opt/checkpoints/hf_llama31_8B_nemo2.nemo \
    --model_type llama \
    --triton_model_name llama \
    --tensor_parallelism_size 1

  5. If the test yields a shared memory-related error, increase the shared memory size using --shm-size (gradually by 50%, for example).

  6. In a separate terminal, access the running container as follows:

    docker exec -it nemo-fw bash

  7. To send a query to the Triton server, run the following script:

    python /opt/Export-Deploy/scripts/deploy/nlp/query.py -mn llama -p "What is the color of a banana?" -mol 5

  8. To export and deploy a different model such as Llama3, Mixtral, or Starcoder, change the model_type in the deploy_triton.py <https://github.com/NVIDIA/NeMo/blob/main/scripts/deploy/nlp/deploy_triton.py>_ script. Please check below to see the list of the model types. For quantized qnemo models, specifying model_type is not necessary as this information is included in the TensorRT-LLM checkpoint.

Use a Script to Deploy NeMo LLMs on a Triton Server#

You can deploy a LLM from a NeMo checkpoint on Triton using the provided script.

Export and Deploy a LLM Model#

After executing the script, it will export the model to TensorRT-LLM and then initiate the service on Triton.

  1. Start the container using the steps described in the Quick Example section.

  2. To begin serving the downloaded model or a quantized model, run the following script:

    python /opt/Export-Deploy/scripts/deploy/nlp/deploy_triton.py \
    --nemo_checkpoint /opt/checkpoints/hf_llama31_8B_nemo2.nemo \
    --model_type llama \
    --triton_model_name llama \
    --tensor_parallelism_size 1

    The following parameters are defined in the deploy_triton.py script:

    • --nemo_checkpoint: path of the .nemo or .qnemo checkpoint file.

    • --model_type: type of the model. choices=[“gptnext”, “gpt”, “llama”, “falcon”, “starcoder”, “mixtral”, “gemma”].

    • --triton_model_name: name of the model on Triton.

    • --triton_model_version: version of the model. Default is 1.

    • --triton_port: port for the Triton server to listen for requests. Default is 8000.

    • --triton_http_address: HTTP address for the Triton server. Default is 0.0.0.0.

    • --triton_model_repository: TensorRT temp folder. Default is /tmp/trt_llm_model_dir/.

    • --tensor_parallelism_size: number of GPUs to split the tensors for tensor parallelism. Default is 1.

    • --pipeline_parallelism_size: number of GPUs to split the model for pipeline parallelism. Default is 1.

    • --dtype: data type of the model on TensorRT-LLM. Default is “bfloat16”. Currently, only “bfloat16” is supported.

    • --max_input_len: maximum input length of the model. Default is 256.

    • --max_output_len: maximum output length of the model. Default is 256.

    • --max_batch_size: maximum batch size of the model. Default is 8.

    • --max_num_tokens: maximum number of tokens. Default is None.

    • --opt_num_tokens: optimum number of tokens. Default is None.

    • --lora_ckpt: a checkpoint list of LoRA weights.

    • --use_lora_plugin: activates the LoRA plugin which enables embedding sharing.

    • --lora_target_modules: adds LoRA to specified modules, but only activates when the --use_lora_plugin is enabled.

    • --max_lora_rank: maximum LoRA rank for different LoRA modules. It is used to compute the workspace size of the LoRA plugin.

    • --no_paged_kv_cache: disables the paged kv cache in the TensorRT-LLM.

    • --disable_remove_input_padding: disables the remove input padding option of TensorRT-LLM.

    • --use_parallel_embedding: enables the parallel embedding feature of TensorRT-LLM.

    • --export_fp8_quantized: manually overrides the FP8 quantization settings.

    • --use_fp8_kv_cache: manually overrides the FP8 KV-cache quantization settings.

    Note: The parameters described here are generalized and should be compatible with any NeMo checkpoint. It is important, however, that you check the LLM model table in the main :ref:Deploy NeMo LLM page <deploy-nemo-framework-models-llm> for optimized inference model compatibility. We are actively working on extending support to other checkpoints.

  3. To export and deploy a different model such as Llama3, Mixtral, or Starcoder, change the model_type in deploy_triton.py <https://github.com/NVIDIA/NeMo/blob/main/scripts/deploy/nlp/deploy_triton.py>_ script. Please see the table below to learn more about which model_type is used for a LLM model (not required for .qnemo models).

    Model Name

    model_type

    GPT

    gpt

    Nemotron

    gpt

    Llama

    llama

    Gemma

    gemma

    StarCoder1

    starcoder

    StarCoder2

    starcoder

    Mistral

    llama

    Mixtral

    mixtral

  4. Whenever the script is executed, it initiates the service by exporting the NeMo checkpoint to the TensorRT-LLM. If you want to skip the exporting step in the optimized inference option, you can specify an empty directory to save the TensorRT-LLM engine produced. Stop the running container and then run the following command to specify an empty directory:

    mkdir tmp_triton_model_repository

docker run --gpus all -it --rm --shm-size=4g -p 8000:8000 \
    -v ${PWD}:/opt/checkpoints/ \
    -w /opt/NeMo \
    nvcr.io/nvidia/nemo:vr

python /opt/Export-Deploy/scripts/deploy/nlp/deploy_triton.py \
    --nemo_checkpoint /opt/checkpoints/hf_llama31_8B_nemo2.nemo \
    --model_type llama \
    --triton_model_name llama \
    --triton_model_repository /opt/checkpoints/tmp_triton_model_repository \
    --tensor_parallelism_size 1

    The checkpoint will be exported to the specified folder after executing the script mentioned above so that it can be reused later.

  5. To load the exported model directly, run the following script within the container:

    python /opt/Export-Deploy/scripts/deploy/nlp/deploy_triton.py \
    --triton_model_name llama \
    --triton_model_repository /opt/checkpoints/tmp_triton_model_repository \
    --model_type llama

  6. Access the models with a Hugging Face token.

    If you want to run inference using the StarCoder1, StarCoder2, or LLama3 models, you’ll need to generate a Hugging Face token that has access to these models. Visit Hugging Face <https://huggingface.co/>__ for more information. After you have the token, perform one of the following steps.

    • Log in to Hugging Face:

    huggingface-cli login

    • Or, set the HF_TOKEN environment variable:

    export HF_TOKEN=your_token_here

Export and Deploy a LLM Model with TensorRT-LLM API#

Alternatively, if the deploy_triton script is unable to export your model to TensorRT-LLM, you can leverage the new TensorRT-LLM LLM API. This API provides a streamlined way to export and deploy models. See the example below:

python /opt/Export-Deploy/scripts/deploy/nlp/deploy_triton.py \
   --nemo_checkpoint /opt/checkpoints/hf_llama31_8B_nemo2.nemo \
   --model_type llama \
   --triton_model_name llama \
   --tensor_parallelism_size 1

After starting the Triton server, you can query the deployed model using the /opt/Export-Deploy/scripts/deploy/nlp/query.py script, as demonstrated in the previous steps.

Use NeMo Export and Deploy Module APIs to Run Inference#

Up until now, we have used scripts for exporting and deploying LLM models. However, NeMo’s deploy and export modules offer straightforward APIs for deploying models to Triton and exporting NeMo checkpoints to TensorRT-LLM.

Export a LLM Model to TensorRT-LLM#

You can use the APIs in the export module to export a NeMo checkpoint to TensorRT-LLM. The following code example assumes the hf_llama31_8B_nemo2.nemo checkpoint has already been downloaded and mounted to the /opt/checkpoints/ path. Additionally, the /opt/checkpoints/tmp_trt_llm path is also assumed to exist.

  1. Run the following command:

    from nemo_export.tensorrt_llm import TensorRTLLM

trt_llm_exporter = TensorRTLLM(model_dir="/opt/checkpoints/tmp_trt_llm/")
trt_llm_exporter.export(
    nemo_checkpoint_path="/opt/checkpoints/hf_llama31_8B_nemo2.nemo",
    model_type="llama",
    tensor_parallelism_size=1,
)

trt_llm_exporter.forward(
    ["What is the best city in the world?"],
    max_output_token=15,
    top_k=1,
    top_p=0.0,
    temperature=1.0,
)

  2. Be sure to check the TensorRTLLM class docstrings for details.

  3. For quantized qnemo models – see Quantization – the same TensorRTLLM class can be used. In this case, specifying model_type is not necessary. Alternatively, advanced users can build engines directly using trtllm-build command, see TensorRT-LLM documentation

    trtllm-build \
    --checkpoint_dir /opt/checkpoints/llama3-70b-base-fp8-qnemo \
    --output_dir /path/to/trt_llm_engine_folder \
    --gemm_plugin fp8 \
    --max_batch_size 8 \
    --max_input_len 2048 \
    --max_seq_len 2560 \
    --workers 2

Deploy a LLM Model to TensorRT-LLM#

You can use the APIs in the deploy module to deploy a TensorRT-LLM model to Triton. The following code example assumes the hf_llama31_8B_nemo2.nemo checkpoint has already been downloaded and mounted to the /opt/checkpoints/ path. Additionally, the /opt/checkpoints/tmp_trt_llm path is also assumed to exist.

  1. Run the following command:

    from nemo_export.tensorrt_llm import TensorRTLLM
from nemo_deploy import DeployPyTriton

trt_llm_exporter = TensorRTLLM(model_dir="/opt/checkpoints/tmp_trt_llm/")
trt_llm_exporter.export(
    nemo_checkpoint_path="/opt/checkpoints/hf_llama31_8B_nemo2.nemo",
    model_type="llama",
    tensor_parallelism_size=1,
)

nm = DeployPyTriton(model=trt_llm_exporter, triton_model_name="llama", http_port=8000)
nm.deploy()
nm.serve()

  2. The same instructions apply to quantized checkpoints.

Direct TensorRT-LLM Export for FP8-trained Models#

If you have a FP8-trained checkpoint, produced during pre-training or fine-tuning with NVIDIA Transformer Engine, you can convert it to a FP8 TensorRT-LLM engine directly using nemo.export. The entry point is the same as with regular .nemo and .qnemo checkpoints:

from nemo_export.tensorrt_llm import TensorRTLLM

trt_llm_exporter = TensorRTLLM(model_dir="/opt/checkpoints/tmp_trt_llm/")
trt_llm_exporter.export(
    nemo_checkpoint_path="/opt/checkpoints/llama2-7b-base-fp8.nemo",
    model_type="llama",
)
trt_llm_exporter.forward(["Hi, how are you?", "I am good, thanks, how about you?"])

The export settings for quantization can be adjusted via trt_llm_exporter.export arguments:

  • fp8_quantized: Optional[bool] = None: enables/disables FP8 quantization

  • fp8_kvcache: Optional[bool] = None: enables/disables FP8 quantization for KV-cache

By default, the quantization settings are auto-detected from the NeMo checkpoint.