lanuch training task:

# [reference]: https://www.youtube.com/watch?v=KaAJtI1T2x4&list=PL_lsbAsL_o2CSuhUhJIiW0IkdT5C2wGWj

simple launch on one node:

python train/main.py  # export CUDA_VISIBLE_DEVICES=0,1  # only use GPU 0 and GPU 1

DDP (FSDP) launch on one node by torchrun (e.g. 8 GPUs):

torchrun --standalone --nproc_per_node=8 train/main.py

DDP (FSDP) launch on multi node by torchrun (e.g. 2 * 8 GPUs, two nodes):

# on node 0#
torchrun --nproc_per_node=8 --nnodes=2 -node_rank=0 --rdzv_backend=c10d --rdzv_endpoint=xxx.xxx.xxx.xxx:xxxx train/main.py

# on node 1#
torchrun --nproc_per_node=8 --nnodes=2 -node_rank=1 --rdzv_backend=c10d --rdzv_endpoint=xxx.xxx.xxx.xxx:xxxx train/main.py

lanuch generation task:

python gen/main.py

lanuch generation task with tensor parallelism:

torchrun --standalone --nproc_per_node=2 gen/main.py

lanuch test code:

python test/module_tests.py

llama3 configs:

llama3_8B

dim=4096, n_layers=32, n_heads=32, n_kv_heads=8, vocab_size=128256,
multiple_of=1024, ffn_dim_multiplier=1.3, norm_eps=1e-5, rope_theta=500000.0,
max_batch_size=32, max_seq_len=2048

llama3_70B

dim=8192, n_layers=80, n_heads=64, n_kv_heads=8, vocab_size=128256,
multiple_of=4096, ffn_dim_multiplier=1.3, norm_eps=1e-5, rope_theta=500000.0,
max_batch_size=32, max_seq_len=2048

llama3_405B

dim=16384, n_layers=126, n_heads=128, n_kv_heads=8, vocab_size=128256,
multiple_of=None, ffn_dim_multiplier=1.3, norm_eps=1e-5, rope_theta=500000.0,
max_batch_size=32, max_seq_len=2048

env configuration:

base pytorch env

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

pip install transformers==4.44.0

pip install tiktoken==0.7.0

pip install blobfile==3.0.0

pip install tqdm==4.66.5

faiss env

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

pip install faiss-cpu==1.8.0.post1

pip install transformers==4.44.0

pip install tiktoken==0.7.0

pip install blobfile==3.0.0

pip install tqdm==4.66.5

fairscale env

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

pip install fairscale==0.4.13

pip install fire==0.7.0

pip install transformers==4.44.0

pip install tiktoken==0.7.0

pip install blobfile==3.0.0

pip install tqdm==4.66.5

tensorrt-llm

cd TensorRT-LLM/examples/bloom

pip install torch torchvision torchaudio (2.4.0, cuda 12.1)

conda install -y mpi4py

conda install openmpi

pip install tensorrt_llm==0.13.0.dev2024081300 --extra-index-
url https://pypi.nvidia.com

pip install -r ./requirements.txt

Quant Things

Pre Doings

Depend On Docker Env

1. Prepare source code

   git lfs install   # Open git large file transport
   git clone https://github.com/NVIDIA/TensorRT-LLM.git
   git checkout v0.8.0  #change version to 0.8.0
   git submodule update --init --recursive  # update some sumbodules
   git lsf pull

2. build docker image

   make -C docker release_build

   and you will get a docker image

3. run docker container

   run from scratch

   make -C docker release_run

   mapping the file is better

   makr -C docker release_run DOCKER_RUN_ARGS="-v  /host/machine/path/:/container/path"
   

   Tips: Run this

   cd This path

   then, do some about quant things 👇，good luck😀

cd /app/tensorrt_llm/examples/llama

Intro

Quantization happened in generate checkpoint stage. Beside convert_checkpoint.py, another script is quantize.py, This two scripts include different quant functions. Here is a demo, shows different quantization methods, LLaMA 7B based.

Int8 weight-only

python convert_checkpoint.py --model_dir /home/Llama-2-7b-hf/  --output_dir ./checkpoint_1gpu_int8_wo --dtype float16 --use_weight_only --weight_only_precision int8

Int4 wight-only

python convert_checkpoint.py --model_dir /home/Llama-2-7b-hf/  --output_dir ./checkpoint_1gpu_int8_wo --dtype float16 --use_weight_only --weight_only_precision int4

Int4 AWQ

python ../quantization/quantize.py --model_dir /home/Llama-2-7b-hf/  --output_dir ./checkpoint_1gpu_int8_wo --dtype float16 --qformat int4_awq --awq_block_size 128 --calib_size 32

Int4 GPTQ

use part of full model model-00001-of-00002.safetensors

python convert_checkpoint.py --model_dir /home/Llama-2-7b-hf/  --output_dir ./checkpoint_1gpu_int8_wo --dtype float16 --use_weight_only --weight_only_precision int4_gptq --per_group --ammo_quant_ckpt_path ./llama-7b-4bit-gs128.safetensors

Int8 SmoothQuant

This quant method also in convert_checkpoint.py, use the --smoothquant parameter to open.

also use --per_token & --per_channel , 0.5 means transfer 50% activation quant stress to weight quant, now let's us consider the corner case, if 0, that mean none-transfer the quant stress, it's hard to quant activation part, easy to quant weight part. if 1,that mean transfer all stress to weight part, cause easy to quant activate part, hard to quant weight part.

python convert_checkpoint.py --model_dir /home/Llama-2-7b-hf/  --output_dir ./checkpoint_1gpu_int8_wo --dtype float16 --smoothquant 0.5 --per_token --per_channel

Tips: if you meet trust_remote_code problem, just run export HF_DATASETS_TRUST_REMOTE_CODE=1 in your terminal, it will download cnn_stories.tgz and dailymail_stories.tgz.

FP8

--calib_size 512 Optional Parameter

python ../quantization/quantize.py --model_dir /home/Llama-2-7b-hf/  --output_dir ./checkpoint_1gpu_int8_wo --dtype float16 --qformat fp8 --calib_size 512

INT8 KV_Cache

python convert_checkpoint.py --model_dir /home/Llama-2-7b-hf/  --output_dir ./checkpoint_1gpu_int8_wo --dtype float16 --int8_kv_cache

👆 file type .safetensors to .safetensors

👇 file type .safetensors to .engine

Engine Generate

After quant, safetensors to engine

trtllm-build --checkpoint_dir ./checkpoint_1gpu_int8_wo --output_dir ./engine_1gpu_int8_wo --gemm_plugin float16 --max_batch_size 16 --max_input_len 2048 --max_output_len 2048

If you meet this problem ValueError: max_workers must be greater than 0, need to mapping CUDA in Container.

sudo make -C docker release_run DOCKER_RUN_ARGS=" --gpus all -v /data_ws/Data_1/tinghao/boxue/llm_weight:/home"

and you see gpus now,

the get the engine😀

Outro

Actually，lots of quant methods in TensorRT-LLM/examples/llama at main · NVIDIA/TensorRT-LLM (github.com). Can test them one by one.

LLama Like Model All Steps

~ From begin to end,from train to quant ~

Intro

All files in myllama, path in docker /home/myllama

First of all, generate and organize files, LLama file type like ! :)

Include config.json,

model-00001-of-00004.safetensors,

model-00002-of-00004.safetensors,

model-00003-of-00004.safetensors,

model-00004-of-00004.safetensors,

pytorch_model-00001-of-00004.bin,

pytorch_model-00002-of-00004.bin,

pytorch_model-00003-of-00004.bin,

pytorch_model-00004-of-00004.bin,

model.safetensors.index.json

fp16 To int8

python convert_checkpoint.py --model_dir /home/Llama-2-7b-hf/ --output_dir ./myllama_1gpu_int8_wo --dtype float16 --use_weight_only --weight_only_precision int8

skip the type check function :)

safetensors To engine

trtllm-build --checkpoint_dir ./myllama_1gpu_int8_wo --output_dir ./myllama_1gpu_int8_wo_engine --gemm_plugin float16 --max_batch_size 16 --max_input_len 2048 --max_output_len 2048

Inference

 python3 run.py --engine_dir ./llama/myllama_1gpu
_int8_wo_engine/ --max_output_len 100 --tokenizer_dir /home/Llama-2-7b-hf/ --input_text "How do I count to nine in French?"

Ref

1. TensorRT-LLM/examples/llama at main · NVIDIA/TensorRT-LLM (github.com)
2. Installing the NVIDIA Container Toolkit — NVIDIA Container Toolkit 1.16.2 documentation

TODO:

1. add AWS s3 support to let several nodes read data from AWS S3 bucket during multi-node training task

some useful links:

quant

https://chatgpt.com/share/31aa8af3-dce2-457f-85db-2b18b3c242ce

torch.distributed

https://pytorch.org/docs/stable/distributed.html The package (torch.distributed) needs to be initialized using the torch.distributed.init_process_group() or torch.distributed.device_mesh.init_device_mesh() function before calling any other methods. Both block until all processes have joined.

torch.distributed.tensor.parallel

https://pytorch.org/docs/stable/distributed.tensor.parallel.html

Pytorch: Pipeline Parallelism

https://pytorch.org/docs/stable/distributed.pipelining.html https://pytorch.org/tutorials/intermediate/pipelining_tutorial.html

huggingface safetensors to llama.cpp gguf format

https://huggingface.co/docs/transformers/main/en/gguf

useful links about Transformer

https://huggingface.co/docs/transformers/main/model_doc/llama2 https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L255

https://huggingface.co/docs/transformers/main/model_doc/llama3 https://github.com/meta-llama/llama3

https://spaces.ac.cn/archives/9708 https://spaces.ac.cn/archives/9948

https://github.com/bojone/rerope?tab=readme-ov-file

https://github.com/hkproj/pytorch-lora

https://github.com/pytorch/torchtitan/tree/main

https://huggingface.co/docs/transformers/main/en/gguf

https://pytorch.org/docs/stable/distributed.pipelining.html https://pytorch.org/tutorials/intermediate/pipelining_tutorial.html