ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

LLM: update Alpaca QLoRA finetuning example on GPU (#9492)

Browse source
This commit is contained in:
binbin Deng 2023-11-21 14:22:19 +08:00 committed by GitHub
parent 566ec85113
commit b7ae572ac3
8 changed files with 139 additions and 30 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
python/llm/example/GPU/QLoRA-FineTuning/alpaca-qlora/README.md
View file

@ -40,6 +40,12 @@ bash finetune_llama2_7b_arc_1_card.sh
bash finetune_llama2_7b_arc_2_card.sh
```
#### Finetuning LLaMA2-7B on single Data Center GPU Flex 170
```bash
bash finetune_llama2_7b_flex_170_1_card.sh
```
#### Finetuning LLaMA2-7B on three Data Center GPU Flex 170
```bash
@ -58,6 +64,18 @@ bash finetune_llama2_7b_pvc_1100_1_card.sh
bash finetune_llama2_7b_pvc_1100_4_card.sh
```
#### Finetuning LLaMA2-7B on single Intel Data Center GPU Max 1550
```bash
bash finetune_llama2_7b_pvc_1550_1_card.sh
```
#### Finetuning LLaMA2-7B on four Intel Data Center GPU Max 1550
```bash
bash finetune_llama2_7b_pvc_1550_4_card.sh
```
### 4. Sample Output
```log
{'loss': 1.9231, 'learning_rate': 2.9999945367033285e-05, 'epoch': 0.0}

15
python/llm/example/GPU/QLoRA-FineTuning/alpaca-qlora/alpaca_qlora_finetuning.py
View file

@ -53,6 +53,21 @@ from bigdl.llm.transformers import AutoModelForCausalLM
# import them from bigdl.llm.transformers.qlora to get a BigDL-LLM compatible Peft model
from bigdl.llm.transformers.qlora import get_peft_model, prepare_model_for_kbit_training
def get_int_from_env(env_keys, default):
"""Returns the first positive env value found in the `env_keys` list or the default."""
for e in env_keys:
val = int(os.environ.get(e, -1))
if val >= 0:
return val
return default
local_rank = get_int_from_env(["LOCAL_RANK","MPI_LOCALRANKID"], "0")
world_size = get_int_from_env(["WORLD_SIZE","PMI_SIZE"], "1")
port = get_int_from_env(["MASTER_PORT"], 29500)
os.environ["LOCAL_RANK"] = str(local_rank)
os.environ["WORLD_SIZE"] = str(world_size)
os.environ["RANK"] = str(local_rank)
os.environ["MASTER_PORT"] = str(port)
def train(
# model/data params

14
python/llm/example/GPU/QLoRA-FineTuning/alpaca-qlora/finetune_llama2_7b_arc_2_card.sh
View file

@ -14,15 +14,13 @@
# limitations under the License.
#
export CCL_ZE_IPC_EXCHANGE=sockets
export MASTER_ADDR=127.0.0.1
export OMP_NUM_THREADS=6 # adjust this to 1/4 of total physical cores
export FI_PROVIDER=tcp
export CCL_ATL_TRANSPORT=ofi
torchrun --standalone \
--nnodes=1 \
--nproc-per-node 2 \
./alpaca_qlora_finetuning.py \
--base_model "meta-llama/Llama-2-7b-hf" \
--data_path "yahma/alpaca-cleaned" \
--output_dir "./bigdl-qlora-alpaca"
mpirun -n 2 \
python -u ./alpaca_qlora_finetuning.py \
--base_model /mnt/disk1/models/Llama-2-7b-hf \
--data_path '/home/arda/binbin/dataset/alpaca-cleaned/alpaca_data_cleaned.json' \
--output_dir "./bigdl-qlora-alpaca" > training.log

23
python/llm/example/GPU/QLoRA-FineTuning/alpaca-qlora/finetune_llama2_7b_flex_170_1_card.sh Normal file
View file

@ -0,0 +1,23 @@
#
# Copyright 2016 The BigDL Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# You could also specify `--base_model` to the local path of the huggingface model checkpoint folder and `--data_path` to the local path of the dataset JSON file
python ./alpaca_qlora_finetuning.py \
--micro_batch_size 2 \
--batch_size 128 \
--base_model "meta-llama/Llama-2-7b-hf" \
--data_path "yahma/alpaca-cleaned" \
--output_dir "./bigdl-qlora-alpaca"

23
python/llm/example/GPU/QLoRA-FineTuning/alpaca-qlora/finetune_llama2_7b_flex_170_3_card.sh
View file

@ -14,17 +14,16 @@
# limitations under the License.
#
export CCL_ZE_IPC_EXCHANGE=sockets
export MASTER_ADDR=127.0.0.1
export OMP_NUM_THREADS=16 # adjust this to 1/4 of total physical cores
export OMP_NUM_THREADS=12 # adjust this to 1/4 of total physical cores
export FI_PROVIDER=tcp
export CCL_ATL_TRANSPORT=ofi
torchrun --standalone \
--nnodes=1 \
--nproc-per-node 3 \
./alpaca_qlora_finetuning.py \
--base_model "meta-llama/Llama-2-7b-hf" \
--data_path "yahma/alpaca-cleaned" \
--output_dir "./bigdl-qlora-alpaca" \
--gradient_checkpointing False \
--micro_batch_size 2 \
--batch_size 128
mpirun -n 3 \
python -u ./alpaca_qlora_finetuning.py \
--base_model "meta-llama/Llama-2-7b-hf" \
--data_path "yahma/alpaca-cleaned" \
--output_dir "./bigdl-qlora-alpaca" \
--gradient_checkpointing False \
--micro_batch_size 2 \
--batch_size 128 > training.log

20
python/llm/example/GPU/QLoRA-FineTuning/alpaca-qlora/finetune_llama2_7b_pvc_1100_4_card.sh
View file

@ -14,15 +14,15 @@
# limitations under the License.
#
export CCL_ZE_IPC_EXCHANGE=sockets
export MASTER_ADDR=127.0.0.1
export OMP_NUM_THREADS=28 # adjust this to 1/4 of total physical cores
torchrun --standalone \
--nnodes=1 \
--nproc-per-node 4 \
./alpaca_qlora_finetuning.py \
--base_model "meta-llama/Llama-2-7b-hf" \
--data_path "yahma/alpaca-cleaned" \
--output_dir "./bigdl-qlora-alpaca" \
--micro_batch_size 8 \
--batch_size 128
export FI_PROVIDER=tcp
export CCL_ATL_TRANSPORT=ofi
mpirun -n 4 \
python -u ./alpaca_qlora_finetuning.py \
--base_model "meta-llama/Llama-2-7b-hf" \
--data_path "yahma/alpaca-cleaned" \
--output_dir "./bigdl-qlora-alpaca" \
--micro_batch_size 8 \
--batch_size 128 > training.log

28
python/llm/example/GPU/QLoRA-FineTuning/alpaca-qlora/finetune_llama2_7b_pvc_1550_1_card.sh Normal file
View file

@ -0,0 +1,28 @@
#
# Copyright 2016 The BigDL Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
export MASTER_ADDR=127.0.0.1
export OMP_NUM_THREADS=28 # adjust this to 1/4 of total physical cores
export FI_PROVIDER=tcp
export CCL_ATL_TRANSPORT=ofi
mpirun -n 2 \
python -u ./alpaca_qlora_finetuning.py \
--base_model "meta-llama/Llama-2-7b-hf" \
--data_path "yahma/alpaca-cleaned" \
--output_dir "./bigdl-qlora-alpaca" \
--micro_batch_size 8 \
--batch_size 128 > training.log

28
python/llm/example/GPU/QLoRA-FineTuning/alpaca-qlora/finetune_llama2_7b_pvc_1550_4_card.sh Normal file
View file

@ -0,0 +1,28 @@
#
# Copyright 2016 The BigDL Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
export MASTER_ADDR=127.0.0.1
export OMP_NUM_THREADS=28 # adjust this to 1/4 of total physical cores
export FI_PROVIDER=tcp
export CCL_ATL_TRANSPORT=ofi
mpirun -n 8 \
python -u ./alpaca_qlora_finetuning.py \
--base_model "meta-llama/Llama-2-7b-hf" \
--data_path "yahma/alpaca-cleaned" \
--output_dir "./bigdl-qlora-alpaca" \
--micro_batch_size 8 \
--batch_size 128 > training.log