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

Replace Llama 7b to Llama2-7b in README.md (#9055)

Browse source 
* Replace Llama 7b with Llama2-7b in README.md

Need to replace the base model to Llama2-7b as we are operating on Llama2 here.

* Replace Llama 7b to Llama2-7b in README.md

a llama 7b in the 1st line is missed

* Update architecture graph

---------

Co-authored-by: Heyang Sun <60865256+Uxito-Ada@users.noreply.github.com>
This commit is contained in:
Ziteng Zhang 2023-09-26 09:56:46 +08:00 committed by GitHub
parent cc84ed70b3
commit a717352c59
1 changed files with 5 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
docker/llm/finetune/lora/README.md
View file

@ -1,16 +1,16 @@
## Run BF16-Optimized Lora Finetuning on Kubernetes with OneCCL ## Run BF16-Optimized Lora Finetuning on Kubernetes with OneCCL
[Alpaca Lora](https://github.com/tloen/alpaca-lora/tree/main) uses [low-rank adaption](https://arxiv.org/pdf/2106.09685.pdf) to speed up the finetuning process of base model [Llama 7b](https://huggingface.co/decapoda-research/llama-7b-hf), and tries to reproduce the standard Alpaca, a general finetuned LLM. This is on top of Hugging Face transformers with Pytorch backend, which natively requires a number of expensive GPU resources and takes significant time. [Alpaca Lora](https://github.com/tloen/alpaca-lora/tree/main) uses [low-rank adaption](https://arxiv.org/pdf/2106.09685.pdf) to speed up the finetuning process of base model [Llama2-7b](https://huggingface.co/meta-llama/Llama-2-7b), and tries to reproduce the standard Alpaca, a general finetuned LLM. This is on top of Hugging Face transformers with Pytorch backend, which natively requires a number of expensive GPU resources and takes significant time.
By constract, BigDL here provides a CPU optimization to accelerate the lora finetuning of Llama 7b, in the power of mixed-precision and distributed training. Detailedly, [Intel OneCCL](https://www.intel.com/content/www/us/en/developer/tools/oneapi/oneccl.html), an available Hugging Face backend, is able to speed up the Pytorch computation with BF16 datatype on CPUs, as well as parallel processing on Kubernetes enabled by [Intel MPI](https://www.intel.com/content/www/us/en/developer/tools/oneapi/mpi-library.html). By constract, BigDL here provides a CPU optimization to accelerate the lora finetuning of Llama2-7b, in the power of mixed-precision and distributed training. Detailedly, [Intel OneCCL](https://www.intel.com/content/www/us/en/developer/tools/oneapi/oneccl.html), an available Hugging Face backend, is able to speed up the Pytorch computation with BF16 datatype on CPUs, as well as parallel processing on Kubernetes enabled by [Intel MPI](https://www.intel.com/content/www/us/en/developer/tools/oneapi/mpi-library.html).
The architecture is illustrated in the following: The architecture is illustrated in the following:
![image](https://github.com/Uxito-Ada/BigDL/assets/60865256/139cf9be-10e6-48df-bc84-8872457e83dd) ![image](https://github.com/Jasonzzt/BigDL/assets/60865256/b66416bc-ad07-49af-8cb0-8967dffb5f58)
As above, BigDL implements its MPI training build on [Kubeflow MPI operator](https://github.com/kubeflow/mpi-operator/tree/master), which encapsulates the deployment as MPIJob CRD, and assists users to handle the construction of a MPI worker cluster on Kubernetes, such as public key distribution, SSH connection, and log collection. As above, BigDL implements its MPI training build on [Kubeflow MPI operator](https://github.com/kubeflow/mpi-operator/tree/master), which encapsulates the deployment as MPIJob CRD, and assists users to handle the construction of a MPI worker cluster on Kubernetes, such as public key distribution, SSH connection, and log collection.
Now, let's go to deploy a Lora finetuning to create a LLM from Llama 7b. Now, let's go to deploy a Lora finetuning to create a LLM from Llama2-7b.
**Note: Please make sure you have already have an available Kubernetes infrastructure and NFS shared storage, and install [Helm CLI](https://helm.sh/docs/helm/helm_install/) for Kubernetes job submission.** **Note: Please make sure you have already have an available Kubernetes infrastructure and NFS shared storage, and install [Helm CLI](https://helm.sh/docs/helm/helm_install/) for Kubernetes job submission.**
@ -22,7 +22,7 @@ Follow [here](https://github.com/kubeflow/mpi-operator/tree/master#installation)
Follow [here](https://github.com/intel-analytics/BigDL/tree/main/docker/llm/finetune/lora/docker#prepare-bigdl-image-for-lora-finetuning) to prepare BigDL Lora Finetuning image in your cluster. Follow [here](https://github.com/intel-analytics/BigDL/tree/main/docker/llm/finetune/lora/docker#prepare-bigdl-image-for-lora-finetuning) to prepare BigDL Lora Finetuning image in your cluster.
As finetuning is from a base model, first download [Llama 7b hf model from the public download site of Hugging Face](https://huggingface.co/decapoda-research/llama-7b-hf/tree/main). Then, download [cleaned alpaca data](https://raw.githubusercontent.com/tloen/alpaca-lora/main/alpaca_data_cleaned_archive.json), which contains all kinds of general knowledge and has already been cleaned. Next, move the downloaded files to a shared directory on your NFS server. As finetuning is from a base model, first download [Llama2-7b model from the public download site of Hugging Face](https://huggingface.co/meta-llama/Llama-2-7b). Then, download [cleaned alpaca data](https://raw.githubusercontent.com/tloen/alpaca-lora/main/alpaca_data_cleaned_archive.json), which contains all kinds of general knowledge and has already been cleaned. Next, move the downloaded files to a shared directory on your NFS server.
### 3. Deploy through Helm Chart ### 3. Deploy through Helm Chart