PTQ enables deploying a model in a low-precision format – FP8, INT4, or INT8 – for efficient serving. Different quantization methods are available including FP8 quantization, INT8 SmoothQuant, and INT4 AWQ.
Model quantization has two primary benefits: reduced model memory requirements and increased inference throughput.
In NeMo, quantization is enabled by the Nvidia AMMO library – a unified algorithmic model optimization & deployment toolkit.
The quantization process consists of the following steps:
Loading a model checkpoint using an appropriate parallelism strategy
Calibrating the model to obtain appropriate algorithm-specific scaling factors
Producing an output directory or .qnemo tarball with model config (json), quantized weights (safetensors) and tokenizer config (yaml).
Loading models requires using an AMMO spec defined in megatron.core.inference.gpt.model_specs.py module. Typically the calibration step is lightweight and uses a small dataset to obtain appropriate statistics for scaling tensors. The output directory produced (or a .qnemo tarball) is ready to be used to build a serving engine with the Nvidia TensorRT-LLM library. The engine build step is also available in NeMo project in nemo.deploy and nemo.export modules.
Quantization algorithm can also be conveniently set to "null" to perform only the weights export step using default precision for TensorRT-LLM deployment. This is useful to obtain baseline performance and accuracy results for comparison.
Example
The example below shows how to quantize the Llama2 70b model into FP8 precision, using tensor parallelism of 8 on a single DGX H100 node. The quantized model is designed for serving using 2 GPUs specified with the export.inference_tensor_parallel parameter.
The script must be launched correctly with the number of processes equal to tensor parallelism. This is achieved with the torchrun command below:
torchrun --nproc-per-node 8 examples/nlp/language_modeling/megatron_llama_quantization.py \
model_file=llama2-70b-base-bf16.nemo \
tensor_model_parallel_size=8 \
pipeline_model_parallel_size=1 \
trainer.num_nodes=1 \
trainer.devices=8 \
trainer.precision=bf16 \
quantization.algorithm=fp8 \
export.decoder_type=llama \
export.inference_tensor_parallel=2 \
model_save=llama2-70b-base-fp8-qnemo
The output directory stores the following files:
llama2-70b-base-fp8-qnemo/
├── config.json
├── rank0.safetensors
├── rank1.safetensors
├── tokenizer.model
└── tokenizer_config.yaml
The TensorRT-LLM engine can be conveniently built and run using TensorRTLLM class available in nemo.export submodule:
from nemo.export import TensorRTLLM
trt_llm_exporter = TensorRTLLM(model_dir="/path/to/trt_llm_engine_folder")
trt_llm_exporter.export(
nemo_checkpoint_path="llama2-70b-base-fp8-qnemo",
model_type="llama",
)
trt_llm_exporter.forward(["Hi, how are you?", "I am good, thanks, how about you?"])
Alternatively, it can also be built directly using trtllm-build command, see TensorRT-LLM documentation:
trtllm-build \
--checkpoint_dir llama2-70b-base-fp8-qnemo \
--output_dir /path/to/trt_llm_engine_folder \
--max_batch_size 8 \
--max_input_len 2048 \
--max_output_len 512 \
--strongly_typed
Known issues
Currently in NeMo, quantizing and building TensorRT-LLM engines is limited to single-node use cases.
The supported and tested model family is Llama2. Quantizing other model types is experimental and may not be fully supported.
Please refer to the following papers for more details on quantization techniques.
Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation, 2020
FP8 Formats for Deep Learning, 2022
SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models, 2022
AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration, 2023