Encoder-Decoder

This document shows how to build and run an Encoder-Decoder (Enc-Dec) model in TensorRT LLM on NVIDIA GPUs.

Table of Contents

Encoder-Decoder

Overview

The TensorRT LLM Enc-Dec implementation can be found in tensorrt_llm/models/enc_dec/model.py. The TensorRT LLM Enc-Dec example code is located in examples/models/core/enc_dec:

trtllm-build to build the TensorRT engine(s) needed to run the Enc-Dec model,
run.py to run the inference on an example input text.
Enc-Dec models can have specific implementations, such as the popular T5 family (T5, mT5, Flan-T5, ByT5), BART family (BART, mBART), and FairSeq family (WMTs). They are now merged into a single convert script:
- convert_checkpoint.py to convert weights from HuggingFace or FairSeq format to TRT-LLM format, and split weights for multi-GPU inference,

Usage

The TensorRT LLM Enc-Dec example code locates at examples/models/core/enc_dec. It takes HuggingFace or FairSeq model name as input, and builds the corresponding TensorRT engines. On each GPU, there will be two TensorRT engines, one for Encoder and one for Decoder.

Encoder-Decoder Model Support

The implementation is designed to support generic encoder-decoder models by abstracting the common and derivative components of different model architectures, such as:

It also supports full Tensor Parallelism (TP), Pipeline Parallelism (PP), and a hybrid of the two. Currently, Fused Multi-Head Attention (FMHA) is not yet enabled for T5 family due to its relative attention design.

In this example, we use T5 (t5-small) and Flan-T5 (google/flan-t5-small) to showcase TRT-LLM support on Enc-Dec models. BART models and FairSeq models can follow very similar steps by just replacing the model name.

Download weights from HuggingFace Transformers

git clone https://huggingface.co/t5-small tmp/hf_models/t5-small
git clone https://huggingface.co/google/flan-t5-small tmp/hf_models/flan-t5-small
git clone https://huggingface.co/facebook/bart-large-cnn tmp/hf_models/bart-large-cnn
git clone https://huggingface.co/facebook/mbart-large-50-many-to-one-mmt tmp/hf_models/mbart-large-50-many-to-one-mmt
git clone https://huggingface.co/google/byt5-small tmp/hf_models/byt5-small

Convert and Split Weights

The convert_checkpoint.py script converts weights from HuggingFace or FairSeq format to TRT-LLM format, and splits weights for multi-GPU inference. --tp_size specifies the number of GPUs for tensor parallelism during inference. Pipeline Parallelism size can be set with --pp_size for distributed inference.

The HuggingFace or Fairseq checkpoints of the enc-dec models mentioned in this Readme are all float32 precision. Use --dtype to set the target inference precision during the weight conversion.

After weight conversion, TensorRT LLM converted weights and model configuration will be saved under <out_dir>/<tpX> directory, which is the --checkpoint_dir input path you should give to the next engine building phase.

Take T5 for example:

# Example: build t5-small using 4-way tensor parallelism on a node with 8 GPUs (but only use 4 of them, for demonstration purpose), BF16, enabling beam search up to width=1.
export MODEL_NAME="t5-small" # or "flan-t5-small"
export MODEL_TYPE="t5"
export INFERENCE_PRECISION="bfloat16"
export TP_SIZE=4
export PP_SIZE=1
export WORLD_SIZE=4
export MAX_BEAM_WIDTH=1
python convert_checkpoint.py --model_type ${MODEL_TYPE} \
                --model_dir tmp/hf_models/${MODEL_NAME} \
                --output_dir tmp/trt_models/${MODEL_NAME}/${INFERENCE_PRECISION} \
                --tp_size ${TP_SIZE} \
                --pp_size ${PP_SIZE} \
                --dtype ${INFERENCE_PRECISION}

Build TensorRT engine(s)

TensorRT LLM builds TensorRT engine(s) with flexible controls on different types of optimizations. Note that these are just examples to demonstrate multi-GPU inference. For small models like T5-small, single GPU is usually sufficient.

After engine building, TensorRT engines will be saved under <out_dir>/<tpX> directory, which is the --engine_dir path you should give to the next engine running phase. It is recommended to have /<Y-gpu> in the output path where Y is number of total GPU ranks in a multi-node, multi-GPU setup, because the same Y number GPUs could be executed with different TP (Tensor Parallelism) and PP (Pipeline Parallelism) combinations.

We should distinguish between X - TP size and Y - total number of GPU ranks:

When X = Y, only TP is enabled
When X < Y, both TP and PP are enabled. In such case, please make sure you have completed weight conversion step for TP=X.

The default value of --max_input_len is 1024. When building DecoderModel, specify decoder input length with --max_input_len=1 for encoder-decoder model to start generation from decoder_start_token_id of length 1. If the start token is a single token (the default behavior of T5/BART/etc.), you should set --max_input_len as 1; if you want the decoder-only type of generation, set --max_input_len above 1 to get similar behavior as HF's decoder_forced_input_ids.

EncoderModel does not generate prompt. --max_seq_len should be the same as --max_input_len. --max_seq_len would be set as --max_input_len if not specified.

DecoderModel takes --max_encoder_input_len and --max_input_len as model inputs, --max_encoder_input_len is set to 1024 as default since --max_input_len is 1024 for EncoderModel.

To be noted:

For T5, add --context_fmha disable. FMHA with T5's relative attention bias is not implemented. Add --use_implicit_relative_attention when --max_seq_len is extremely large, causing decoder engine size to be too large to fit in memory. Compute relative attention on-the-fly (implicitly, without pre-computation) instead.
--bert_attention_plugin, --gpt_attention_plugin, --remove_input_padding, --gemm_plugin require explicit disabling and setting, or else they'll be set to default value in trtllm-build.

# --gpt_attention_plugin is necessary in Enc-Dec.
# Try --gemm_plugin to prevent accuracy issue.
# It is recommended to use --remove_input_padding along with --gpt_attention_plugin for better performance
trtllm-build --checkpoint_dir tmp/trt_models/${MODEL_NAME}/${INFERENCE_PRECISION}/encoder \
                --output_dir tmp/trt_engines/${MODEL_NAME}/${INFERENCE_PRECISION}/encoder \
                --paged_kv_cache disable \
                --moe_plugin disable \
                --max_beam_width ${MAX_BEAM_WIDTH} \
                --max_batch_size 8 \
                --max_input_len 1024 \
                --gemm_plugin ${INFERENCE_PRECISION} \
                --bert_attention_plugin ${INFERENCE_PRECISION} \
                --gpt_attention_plugin ${INFERENCE_PRECISION} \
                --remove_input_padding enable \
                --context_fmha disable

# For decoder, refer to the above content and set --max_input_len correctly
trtllm-build --checkpoint_dir tmp/trt_models/${MODEL_NAME}/${INFERENCE_PRECISION}/decoder \
                --output_dir tmp/trt_engines/${MODEL_NAME}/${INFERENCE_PRECISION}/decoder \
                --moe_plugin disable \
                --max_beam_width ${MAX_BEAM_WIDTH} \
                --max_batch_size 8 \
                --max_input_len 1 \
                --max_seq_len 201 \
                --max_encoder_input_len 1024 \
                --gemm_plugin ${INFERENCE_PRECISION} \
                --bert_attention_plugin ${INFERENCE_PRECISION} \
                --gpt_attention_plugin ${INFERENCE_PRECISION} \
                --remove_input_padding enable \
                --context_fmha disable

For BART, --context_fmha can be enabled. trtllm-build has the default setting to enable it.

# Example: build bart-large-cnn using a single GPU, FP32, running greedy search
export MODEL_NAME="bart-large-cnn" # or "mbart-large-50-many-to-one-mmt"
export MODEL_TYPE="bart"
export INFERENCE_PRECISION="float32"
export TP_SIZE=1
export PP_SIZE=1
export WORLD_SIZE=1
export MAX_BEAM_WIDTH=1
python convert_checkpoint.py --model_type ${MODEL_TYPE} \
                --model_dir tmp/hf_models/${MODEL_NAME} \
                --output_dir tmp/trt_models/${MODEL_NAME}/${INFERENCE_PRECISION} \
                --tp_size ${TP_SIZE} \
                --pp_size ${PP_SIZE} \
                --dtype ${INFERENCE_PRECISION}

# Note: non-T5 models can enable FMHA for the encoder part, for FP16/BF16, the default is enabled
trtllm-build --checkpoint_dir tmp/trt_models/${MODEL_NAME}/${INFERENCE_PRECISION}/encoder \
                --output_dir tmp/trt_engines/${MODEL_NAME}/${INFERENCE_PRECISION}/encoder \
                --paged_kv_cache disable \
                --moe_plugin disable \
                --max_beam_width ${MAX_BEAM_WIDTH} \
                --max_batch_size 8 \
                --max_input_len 1024 \
                --gemm_plugin ${INFERENCE_PRECISION} \
                --bert_attention_plugin ${INFERENCE_PRECISION} \
                --gpt_attention_plugin ${INFERENCE_PRECISION} \
                --remove_input_padding enable
                # --context_fmha disable should be removed

# Use the same command for decoder engine
trtllm-build --checkpoint_dir tmp/trt_models/${MODEL_NAME}/${INFERENCE_PRECISION}/decoder \
                --output_dir tmp/trt_engines/${MODEL_NAME}/${INFERENCE_PRECISION}/decoder \
                --moe_plugin disable \
                --max_beam_width ${MAX_BEAM_WIDTH} \
                --max_batch_size 8 \
                --max_input_len 1 \
                --max_seq_len 201 \
                --max_encoder_input_len 1024 \
                --gemm_plugin ${INFERENCE_PRECISION} \
                --bert_attention_plugin ${INFERENCE_PRECISION} \
                --gpt_attention_plugin ${INFERENCE_PRECISION} \
                --remove_input_padding enable
                # --context_fmha disable should be removed

Run

Run a TensorRT LLM Enc-Dec model using the engines generated by build.py. Note that during model deployment, only the TensorRT engine files are needed. Previously downloaded model checkpoints and converted weights can be removed.

Different types of runtime are provided for encoder-decoder models. Following an order of serving performance and good usability, we recommend:

(NEW) Python binding of C++ runtime w/ Paged KV Cache and Inflight Batching (IFB)
Python runtime w/ Static Batching
(NEW) C++ runtime w/ Paged KV Cache and Inflight Batching

Please refer to the documentation for the details of paged kv cache and inflight batching.

Run C++ runtime

Note: to use inflight batching and paged kv cache features in C++ runtime, please make sure you have set --paged_kv_cache enable (which is by default enabled) in the trtllm-build command of the decoder. Meanwhile, if using Python runtime, it is recommended to disable this flag by --paged_kv_cache disable to avoid any unnecessary overhead.

Note that for C++ runtime and Triton backend, Pipeline Parallelism (PP) is not supported yet, because PP usage is relatively rare for encoder-decoder models. If PP is really needed, it is recommended to use the Python runtime instead.

For good usability, Python binding of the C++ runtime is provided. You can use the high-level C++ ModelRunner under the examples/ root folder.

# Inferencing via python binding of C++ runtime with inflight batching (IFB)
python3 ../../../run.py --engine_dir tmp/trt_engines/${MODEL_NAME}/${INFERENCE_PRECISION} --tokenizer_dir tmp/hf_models/${MODEL_NAME} --max_output_len 64 --num_beams=1 --input_text "translate English to German: The house is wonderful."

You can specify --kv_cache_free_gpu_memory_fraction to control the percentage of free GPU memory to be used by KV cache (by default 0.9), and --cross_kv_cache_fraction to control the percentage of KV cache to be used by cross attention (by default 0.5, and rest of the KV cache will be used by self attention).

For pure C++ runtime, there is no example given yet. Please check the Executor API to implement your own end-to-end workflow. It is highly recommended to leverage more encapsulated solutions such as the above C++ Python binding or Triton backend.

Run with Triton Backend

Triton backend contains the tutorial on how to run encoder-decoder engines with Tritonserver.

Run Python runtime

For pure Python runtime, you can still use the encoder-decoder specific script under examples/models/core/enc_dec/.

# Inferencing w/ single GPU greedy search, compare results with HuggingFace FP32
python3 run.py --engine_dir tmp/trt_engines/${MODEL_NAME}/${INFERENCE_PRECISION} --engine_name ${MODEL_NAME} --model_name tmp/hf_models/${MODEL_NAME} --max_new_token=64 --num_beams=1 --compare_hf_fp32

# Inferencing w/ 4 GPUs (4-way TP, as configured during the engine building step), greedy search, compare results with HuggingFace FP32
mpirun --allow-run-as-root -np ${WORLD_SIZE} python3 run.py --engine_dir tmp/trt_engines/${MODEL_NAME}/${INFERENCE_PRECISION} --engine_name ${MODEL_NAME} --model_name tmp/hf_models/${MODEL_NAME} --max_new_token=64 --num_beams=1 --compare_hf_fp32

Benchmark

Benchmark C++ runtime

The tutorial for encoder-decoder C++ runtime benchmark can be found in benchmarks/cpp

Run BART with LoRA

Download the base model and lora model from HF:

git clone https://huggingface.co/facebook/bart-large-cnn tmp/hf_models/bart-large-cnn
git clone https://huggingface.co/sooolee/bart-large-cnn-samsum-lora tmp/hf_models/bart-large-cnn-samsum-lora

If using customize models, just put both the base model and lora model dirs into tmp/hf_models.

Convert and Split Weights, setting --hf_lora_dir.

export INFERENCE_PRECISION="float16"
python convert_checkpoint.py --model_type bart \
                --model_dir tmp/hf_models/bart-large-cnn \
                --output_dir tmp/trt_models/bart-large-cnn/${INFERENCE_PRECISION} \
                --tp_size 1 \
                --pp_size 1 \
                --dtype ${INFERENCE_PRECISION}

Build engine, setting --use_lora_plugin.

trtllm-build --checkpoint_dir tmp/trt_models/bart-large-cnn/${INFERENCE_PRECISION}/encoder \
                --output_dir tmp/trt_engines/bart-large-cnn/${INFERENCE_PRECISION}/encoder \
                --paged_kv_cache disable \
                --moe_plugin disable \
                --max_beam_width 1 \
                --max_batch_size 8 \
                --max_input_len 1024 \
                --gemm_plugin ${INFERENCE_PRECISION} \
                --bert_attention_plugin ${INFERENCE_PRECISION} \
                --gpt_attention_plugin ${INFERENCE_PRECISION} \
                --remove_input_padding disable \
                --lora_plugin ${INFERENCE_PRECISION} \
                --lora_dir tmp/hf_models/bart-large-cnn-samsum-lora/ \
                --lora_target_modules attn_q attn_v

trtllm-build --checkpoint_dir tmp/trt_models/bart-large-cnn/${INFERENCE_PRECISION}/decoder \
                --output_dir tmp/trt_engines/bart-large-cnn/${INFERENCE_PRECISION}/decoder \
                --moe_plugin disable \
                --max_beam_width 1 \
                --max_batch_size 8 \
                --max_input_len 1 \
                --max_seq_len 201 \
                --max_encoder_input_len 1024 \
                --gemm_plugin ${INFERENCE_PRECISION} \
                --bert_attention_plugin ${INFERENCE_PRECISION} \
                --gpt_attention_plugin ${INFERENCE_PRECISION} \
                --remove_input_padding disable \
                --lora_plugin ${INFERENCE_PRECISION} \
                --lora_dir tmp/hf_models/bart-large-cnn-samsum-lora/ \
                --lora_target_modules attn_q cross_attn_q attn_v cross_attn_v

Run the engine, setting --lora_dir and --lora_task_uids. --lora_task_uids should be set as a list of uids which length equals to batch size. The following example is for batch size = 3:

python run.py \
        --engine_dir tmp/trt_engines/bart-large-cnn/${INFERENCE_PRECISION}/ \
        --engine_name bart-large-cnn \
        --model_name tmp/hf_models/bart-large-cnn \
        --max_new_token=64 \
        --num_beams=1 \
        --lora_dir tmp/hf_models/bart-large-cnn-samsum-lora/ \
        --lora_task_uids 0 0 0

Run with multi-loRA, append --lora_dir with other lora directories and set --lora_task_uids according to the index of the lora directories. Set to "-1" to run with the base model:

python run.py \
        --engine_dir tmp/trt_engines/bart-large-cnn/${INFERENCE_PRECISION}/ \
        --engine_name bart-large-cnn \
        --model_name tmp/hf_models/bart-large-cnn \
        --max_new_token=64 \
        --num_beams=1 \
        --lora_dir tmp/hf_models/bart-large-cnn-samsum-lora/ ... \
        --lora_task_uids 0 -1 -1 0 0 -1

Reminders

Flan-T5 models have known issues regarding FP16 precision and using BF16 precision is recommended, regardless of TRT-LLM. Please stay with FP32 or BF16 precision for Flan-T5 family.
For T5 and Flan-T5 family that have relative attention bias design, the relative attention table is split along num_heads dimension in Tensor Parallelism mode. Therefore, num_heads must be divisible by tp_size. Please be aware of this when setting the TP parameter.
For mBART, models that can control output languages (e.g. mbart-large-50-many-to-many-mmt) are not currently supported, as the script does not support ForcedBOSTokenLogitsProcessor to control output languages.

Attention Scaling Factors

The q_scaling convention in the TRT-LLM plugin is defined as follows:

norm_factor = 1.f / (q_scaling * sqrt(head_size))

In the Multi-Head Attention (MHA) mechanism, the output of the Q*K^T product is scaled by this constant value norm_factor as norm_factor * (Q*K^T) for softmax. This scaling factor can be adjusted or neutralized based on the model's requirements.

Handling in Different Models:

BART/FairSeq NMT: For the BART models, q_scaling is set to 1.f. Therefore, the norm_factor for BART becomes 1.f / sqrt(head_size). TRT-LLM uses the default value q_scaling = 1.f. Similar to FairSeq NMT models.
T5: For the T5 models, q_scaling is 1.f/sqrt(head_size), leading to a norm_factor of 1.f. This is handled in T5 by the TRT-LLM's get_offset_q_scaling() function, which reads head_size from the T5 model configuration and sets q_scaling = 1.f/sqrt(head_size) to effectively offset the norm_factor to 1.f.

Run FairSeq NMT (Neural Machine Translation) models

FairSeq model download and library dependency are different from HuggingFace ones. Especially if you are following the recommended docker container setup in README, it has a custom PyTorch build but FairSeq installation will force upgrade the PyTorch version. As a workaround, we skip the torch and torchaudio dependencies in FairSeq to make everything work nicely inside the TRT-LLM container.

# Download weights from HuggingFace Transformers
# Instructions from: https://github.com/facebookresearch/fairseq/blob/main/examples/translation/README.md#example-usage-cli-tools. Public model checkpoints are also listed there. Here we use WMT'14 Transformer model as an example.
mkdir -p tmp/fairseq_models && curl https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-fr.joined-dict.transformer.tar.bz2 | tar xvjf - -C tmp/fairseq_models  --one-top-level=wmt14 --strip-components 1 --no-same-owner

# Install FairSeq dependency
# avoid base torch to be upgraded by fairseq
pushd tmp && (git clone https://github.com/facebookresearch/fairseq.git || true) && pushd fairseq && sed -i '/torch>=/d;/torchaudio>=/d' setup.py && pip install -e . && pip install sacremoses subword_nmt && popd && popd

# Convert and Split Weights, single GPU example
export TP_SIZE=1
export PP_SIZE=1
export WORLD_SIZE=1
export INFERENCE_PRECISION="float32"
python convert_checkpoint.py --model_type nmt \
                --model_dir tmp/fairseq_models/wmt14 \
                --output_dir tmp/trt_models/wmt14/${INFERENCE_PRECISION} \
                --tp_size ${TP_SIZE} \
                --pp_size ${PP_SIZE} \
                --dtype ${INFERENCE_PRECISION}

# Build TensorRT engine(s)
# Note: non-T5 models can enable FMHA for the encoder part, although only FP16/BF16 precisions are valid
trtllm-build --checkpoint_dir tmp/trt_models/wmt14/${INFERENCE_PRECISION}/encoder \
                --output_dir tmp/trt_engines/wmt14/${INFERENCE_PRECISION}/encoder \
                --paged_kv_cache disable \
                --moe_plugin disable \
                --max_beam_width 1 \
                --max_batch_size 8 \
                --max_input_len 1024 \
                --bert_attention_plugin ${INFERENCE_PRECISION} \
                --gpt_attention_plugin ${INFERENCE_PRECISION} \
                --remove_input_padding disable

trtllm-build --checkpoint_dir tmp/trt_models/wmt14/${INFERENCE_PRECISION}/decoder \
                --output_dir tmp/trt_engines/wmt14/${INFERENCE_PRECISION}/decoder \
                --moe_plugin disable \
                --max_beam_width 1 \
                --max_batch_size 8 \
                --max_input_len 1 \
                --max_seq_len 201 \
                --max_encoder_input_len 1024 \
                --bert_attention_plugin ${INFERENCE_PRECISION} \
                --gpt_attention_plugin ${INFERENCE_PRECISION} \
                --remove_input_padding disable
# Run
mpirun --allow-run-as-root -np ${WORLD_SIZE} python3 run.py --engine_dir tmp/trt_engines/wmt14/${INFERENCE_PRECISION} --engine_name wmt14 --model_name tmp/fairseq_models/wmt14/${INFERENCE_PRECISION} --max_new_token=24 --num_beams=1

FP8 Post-Training Quantization

The examples below uses the NVIDIA Modelopt (AlgorithMic Model Optimization) toolkit for the model quantization process.

First make sure Modelopt toolkit nvidia-modelopt>=0.22.1 is installed (see examples/quantization/README.md).

Note

Modelopt 0.22.1 is not yet released.

Get quantized checkpoint with ModelOpt

Currently supported conversion are bart-large-cnn and T5 family. For bart, please set --dtype float16; for T5 family, please set --dtype float32 due to known bug with apex+HF mentioned in transformer:issue/34264.

# Example: quantize bart-large-cnn using 4-way tensor parallelism on a node with 8 GPUs (but only use 4 of them, for demonstration purpose) into FP8 weight and convert to TRTLLM checkpoint.
export MODEL_NAME="bart-large-cnn"
export MODEL_TYPE="bart"
export INFERENCE_PRECISION="float16"
export TP_SIZE=4
export PP_SIZE=1
export WORLD_SIZE=4
export MAX_BEAM_WIDTH=1
python ../quantization/quantize.py \
                --model_dir tmp/hf_models/${MODEL_NAME} \
                --dtype ${INFERENCE_PRECISION} \
                --qformat fp8 \
                --kv_cache_dtype fp8 \
                --output_dir tmp/trt_models/${MODEL_NAME}/fp8 \
                --calib_size 512 \
                --batch_size 16 \
                --tp_size ${TP_SIZE} \
                --pp_size ${PP_SIZE}

The rest may follow the same command in Build TensorRT engine(s), with some notes:

For bart, please add --use_fp8_context_fmha enable for fp8 context fmha support. For t5, context fmha is not supported due to relative attention bias.
Please ensure --paged_kv_cache enable for decoder for fp8 paged kv cache.
Please use --gemm_plugin auto, --bert_attention_plugin auto, --gpt_attention_plugin auto instead of setting precision to these plugins.
Please use CPP runtime for better performance.