Skip to content

Add new example to fine tune llama-2 70b with lora #80

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
lu-wang-dl wants to merge 8 commits into
base: master
Choose a base branch
from
Open

Add new example to fine tune llama-2 70b with lora #80

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
1f7f497
Add example notebooks for mistral model
Oct 5, 2023
4d6870e
fix
Oct 5, 2023
f59abd4
Update and address the comments
Oct 9, 2023
3e7d8ba
address the comments
Oct 9, 2023
cf8eba3
add readme
lu-wang-dl Oct 9, 2023
d4d7fc0
Add example to fine tune llama-2 70B with lora
Oct 13, 2023
f941387
Update llm-models/llamav2/llamav2-70b/07_fine_tune_lora.py
lu-wang-dl Oct 20, 2023
b5e81bc
Update fine_tune_lora.py
lu-wang-dl Oct 20, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,7 @@ table th:nth-of-type(4) {

| Use case | Quality-optimized | Balanced | Speed-optimized |
|----------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------|
| Text generation following instructions | [MPT-30B-Instruct](llm-models/mpt/mpt-30b/) <br> <br> [Llama-2-70b-chat-hf](llm-models/llamav2/llamav2-70b) | [MPT-7B-Instruct](llm-models/mpt/mpt-7b) <br> [MPT-7B-8k-Instruct](llm-models/mpt/mpt-7b-8k) <br> <br> [Llama-2-7b-chat-hf](llm-models/llamav2/llamav2-7b) <br> [Llama-2-13b-chat-hf](llm-models/llamav2/llamav2-13b) | |
| Text generation following instructions | [MPT-30B-Instruct](llm-models/mpt/mpt-30b/) <br> <br> [Llama-2-70b-chat-hf](llm-models/llamav2/llamav2-70b) | [mistral-7b](llm-models/mistral/mistral-7b) <br><br> [MPT-7B-Instruct](llm-models/mpt/mpt-7b) <br> [MPT-7B-8k-Instruct](llm-models/mpt/mpt-7b-8k) <br> <br> [Llama-2-7b-chat-hf](llm-models/llamav2/llamav2-7b) <br> [Llama-2-13b-chat-hf](llm-models/llamav2/llamav2-13b) | |
| Text embeddings (English only) | | [bge-large-en(0.3B)](llm-models/embedding/bge/bge-large) <br> [e5-large-v2 (0.3B)](llm-models/embedding/e5-v2) <br> [instructor-xl (1.3B)](llm-models/embedding/instructor-xl)* | [bge-base-en (0.1B)](llm-models/embedding/bge) <br> [e5-base-v2 (0.1B)](llm-models/embedding/e5-v2) |
| Transcription (speech to text) | | [whisper-large-v2](llm-models/transcription/whisper)(1.6B) <br> [whisper-medium](llm-models/transcription/whisper) (0.8B) | |
| Image generation | | [stable-diffusion-xl](llm-models/image_generation/stable_diffusion) | |
Expand Down
47 changes: 47 additions & 0 deletions llm-models/config/a10_config_zero2.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,47 @@
{
"fp16": {
"enabled": false
},
"bf16": {
"enabled": true
},
"optimizer": {
"type": "AdamW",
"params": {
"lr": "auto",
"betas": "auto",
"eps": "auto",
"weight_decay": "auto"
}
},
"scheduler": {
"type": "WarmupLR",
"params": {
"warmup_min_lr": "auto",
"warmup_max_lr": "auto",
"warmup_num_steps": "auto"
}
},
"zero_optimization": {
"stage": 2,
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 5e7,
"reduce_bucket_size": "auto",
"reduce_scatter": true,
"offload_param": {
"device": "cpu",
"pin_memory": true
},
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
}
},
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"steps_per_print": 50,
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"wall_clock_breakdown": false
}
74 changes: 74 additions & 0 deletions llm-models/llamav2/llamav2-70b/07_fine_tune_lora.py
View file
Open in desktop
Copy link
Contributor

@es94129 es94129 Oct 13, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Since 07 is used for AI gateway, maybe other indices

Copy link
Collaborator Author

@lu-wang-dl lu-wang-dl Oct 20, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure. Let's design a proper orders after.

Original file line number Diff line number Diff line change
@@ -0,0 +1,74 @@
# Databricks notebook source
# MAGIC %md
# MAGIC
# MAGIC # Fine tune llama-2-70b with LoRA and deepspeed on a single node
# MAGIC
# MAGIC [Llama 2](https://huggingface.co/meta-llama) is a collection of pretrained and fine-tuned generative text models ranging in scale from 7 billion to 70 billion parameters. It is trained with 2T tokens and supports context length window upto 4K tokens. [Llama-2-70b-hf](https://huggingface.co/meta-llama/Llama-2-70b-hf) is the 7B pretrained model, converted for the Hugging Face Transformers format.
Copy link
Contributor

@es94129 es94129 Oct 13, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit

Suggested change
# MAGIC [Llama 2](https://huggingface.co/meta-llama) is a collection of pretrained and fine-tuned generative text models ranging in scale from 7 billion to 70 billion parameters. It is trained with 2T tokens and supports context length window upto 4K tokens. [Llama-2-70b-hf](https://huggingface.co/meta-llama/Llama-2-70b-hf) is the 7B pretrained model, converted for the Hugging Face Transformers format.
# MAGIC [Llama 2](https://huggingface.co/meta-llama) is a collection of pretrained and fine-tuned generative text models ranging in scale from 7 billion to 70 billion parameters. It is trained with 2T tokens and supports context length window upto 4K tokens. [Llama-2-70b-hf](https://huggingface.co/meta-llama/Llama-2-70b-hf) is the 70B pretrained model, converted for the Hugging Face Transformers format.

Copy link
Collaborator Author

@lu-wang-dl lu-wang-dl Oct 20, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

deepspeed is used for multi-GPU training with LORA.

# MAGIC
# MAGIC This is to fine-tune [llama-2-70b-hf](https://huggingface.co/meta-llama/Llama-2-70b-hf) models on the [dolly_hhrlhf](https://huggingface.co/datasets/mosaicml/dolly_hhrlhf) dataset.
# MAGIC
# MAGIC Environment for this notebook:
# MAGIC - Runtime: 14.0 GPU ML Runtime
# MAGIC - Instance: `Standard_NC48ads_A100_v4` on Azure with 2 A100-80GB GPUs, `p4d.24xlarge` on AWS with 8 A100-40GB GPUs
# MAGIC
# MAGIC Requirements:
# MAGIC - To get the access of the model on HuggingFace, please visit the [Meta website](https://ai.meta.com/resources/models-and-libraries/llama-downloads) and accept our license terms and acceptable use policy before submitting this form. Requests will be processed in 1-2 days.
# MAGIC

# COMMAND ----------

# MAGIC %md
# MAGIC Install the missing libraries

# COMMAND ----------

# MAGIC %pip install deepspeed==0.9.5 xformers
# MAGIC %pip install git+https://github.com/huggingface/peft.git
# MAGIC %pip install bitsandbytes==0.40.1 einops==0.6.1 trl==0.4.7
# MAGIC %pip install -U torch==2.0.1 accelerate==0.21.0 transformers==4.31.0
# MAGIC dbutils.library.restartPython()

# COMMAND ----------

import os
os.environ["HF_HOME"] = "/local_disk0/hf"
os.environ["HF_DATASETS_CACHE"] = "/local_disk0/hf"
os.environ["TRANSFORMERS_CACHE"] = "/local_disk0/hf"

# COMMAND ----------

from huggingface_hub import notebook_login

# Login to Huggingface to get access to the model
notebook_login()

# COMMAND ----------

# MAGIC %md
# MAGIC ## Fine tune the model with `deepspeed`
# MAGIC
# MAGIC The fine tune logic is written in `scripts/fine_tune_deepspeed.py`. The dataset used for fine tune is [databricks-dolly-15k ](https://huggingface.co/datasets/databricks/databricks-dolly-15k) dataset.
# MAGIC
# MAGIC

# COMMAND ----------

# MAGIC %sh
# MAGIC deepspeed \
# MAGIC --num_gpus 2 \
Copy link
Contributor

@es94129 es94129 Oct 13, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

--num_gpus is probably not needed because deepspeed can use all the GPUs on the machine

Copy link
Collaborator Author

@lu-wang-dl lu-wang-dl Oct 20, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good point. Let me remove it.

# MAGIC scripts/fine_tune_lora.py \
# MAGIC --output_dir="/local_disk0/output"
Copy link
Contributor

@es94129 es94129 Oct 13, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Q: What is the difference between --output_dir and /local_disk0/final_model, is the latter just the LoRA weights?


# COMMAND ----------

# MAGIC %md
# MAGIC Model checkpoint is saved at `/local_disk0/final_model`.

# COMMAND ----------

# MAGIC %sh
# MAGIC ls /local_disk0/final_model
Copy link
Contributor

@es94129 es94129 Oct 13, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you also add instructions or code for how to load this for inference?


# COMMAND ----------


Loading