Files
2025-08-06 15:49:14 +08:00

132 lines
5.3 KiB
Markdown

# BLOOM
This document shows how to build and run a BLOOM model in XTRT-LLM on both single XPU and single node multi-XPU.
## Overview
The XTRT-LLM BLOOM example code is located in [`examples/bloom`](./). There are several main files in that folder:
* [`build.py`](./build.py) to build the XTRT engine(s) needed to run the BLOOM model,
* [`run.py`](./run.py) to run the inference on an input text,
* [`summarize.py`](./summarize.py) to summarize the articles in the [cnn_dailymail](https://huggingface.co/datasets/cnn_dailymail) dataset using the model.
## Support Matrix
* FP16
* INT8 & INT4 Weight-Only
* Tensor Parallel
## Usage
The XTRT-LLM BLOOM example code locates at [examples/bloom](./). It takes HF weights as input, and builds the corresponding XTRT engines. The number of XTRT engines depends on the number of XPUs used to run inference.
### Build XTRT engine(s)
Need to prepare the HF BLOOM checkpoint first by following the guides here https://huggingface.co/docs/transformers/main/en/model_doc/bloom.
e.g. To install BLOOM-560M
```bash
# Setup git-lfs
git lfs install
rm -rf ./downloads/bloom/560M/
mkdir -p ./downloads/bloom/560M/ && git clone https://huggingface.co/bigscience/bloom-560m ./downloads/bloom/560M/
```
XTRT-LLM BLOOM builds XTRT engine(s) from HF checkpoint.
Normally `build.py` only requires single XPU, but if you've already got all the XPUs needed for inference, you could enable parallel building to make the engine building process faster by adding `--parallel_build` argument. Please note that currently `parallel_build` feature only supports single node.
Here're some examples:
```bash
# Build a single-XPU float16 engine from HF weights.
# Try use_gemm_plugin to prevent accuracy issue. TODO check this holds for BLOOM
# Single XPU on BLOOM 560M
python build.py --model_dir ./downloads/bloom/560M/ \
--dtype float16 \
--use_gpt_attention_plugin float16 \
--output_dir ./downloads/bloom/560M/trt_engines/fp16/1-XPU/
# Build the BLOOM 560M using a single XPU and apply INT8 weight-only quantization.
python build.py --model_dir ./downloads/bloom/560M/ \
--dtype float16 \
--use_gpt_attention_plugin float16 \
--use_weight_only \
--weight_only_precision int8 \
--output_dir ./downloads/bloom/560M/trt_engines/int8_weight_only/1-XPU/
# Use 2-way tensor parallelism on BLOOM 560M
python build.py --model_dir ./downloads/bloom/560M/ \
--dtype float16 \
--use_gpt_attention_plugin float16 \
--output_dir ./downloads/bloom/560M/trt_engines/fp16/2-XPU/ \
--world_size 2
```
#### SmoothQuant
Unlike the FP16 build where the HF weights are processed and loaded into the XTRT-LLM directly, the SmoothQuant needs to load INT8 weights which should be pre-processed before building an engine.
Example:
```bash
python3 hf_bloom_convert.py -i ./downloads/bloom/560M/ -o ./downloads/bloom-smooth/560M --smoothquant 0.5 --tensor-parallelism 1 --storage-type float16
```
Note `hf_bloom_convert.py` run with pytorch, and
1. `torch-cpu` has better accuracy than XPyTorch generally.
2. XPyTorch often use more than 32GB GM, thus more XPU are necessary to finish it.
3. add `-p=1` if run with XPyTorch.
[`build.py`](./build.py) add new options for the support of INT8 inference of SmoothQuant models.
`--use_smooth_quant` is the starting point of INT8 inference. By default, it
will run the model in the _per-tensor_ mode.
`--per-token` and `--per-channel` are not supported yet.
Examples of build invocations:
```bash
# Build model for SmoothQuant in the _per_tensor_ mode.
python3 build.py --bin_model_dir=./downloads/bloom-smooth/560M/1-XPU \
--use_smooth_quant \
--use_gpt_attention_plugin float16 \
--output_dir ./downloads/bloom-smooth/560M/trt_engines/fp16/1-XPU/
```
Note that GPT attention plugin is required to be enabled for SmoothQuant for now.
Note we use `--bin_model_dir` instead of `--model_dir` since SmoothQuant model needs INT8 weights and various scales from the binary files.
### Run
```bash
python ../summarize.py --test_trt_llm \
--hf_model_dir ./downloads/bloom/560M/ \
--data_type fp16 \
--engine_dir ./downloads/bloom/560M/trt_engines/fp16/1-XPU/
python ../summarize.py --test_trt_llm \
--hf_model_dir ./downloads/bloom/560M/ \
--data_type fp16 \
--engine_dir ./downloads/bloom/560M/trt_engines/int8_weight_only/1-XPU/
python run.py --tokenizer_dir ./downloads/bloom/560M/ \
--max_output_len=50 \
--engine_dir ./downloads/bloom/560M/trt_engines/fp16/1-XPU/
python run.py --tokenizer_dir ./downloads/bloom/560M/ \
--max_output_len=50 \
--engine_dir ./downloads/bloom/560M/trt_engines/int8_weight_only/1-XPU/
python run.py --tokenizer_dir ./downloads/bloom/560M/ \
--max_output_len=50 \
--engine_dir ./downloads/bloom-smooth/560M/trt_engines/fp16/1-XPU/
mpirun -n 2 --allow-run-as-root \
python run.py --tokenizer_dir ./downloads/bloom/560M/ \
--max_output_len=50 \
--engine_dir ./downloads/bloom/560M/trt_engines/fp16/2-XPU/
```