[Examples] PyTorch (train and inference), TGI and TEI (inference)

[alvarobartt]

As part of our collaboration with Google Cloud and following up #2 to create the Deep Learning Containers (DLCs) for both Google Kubernetes Engine (GKE) and Vertex AI, we want to create dedicated examples per each alternative offered.

The examples to be created and included within this repository are listed below and divided in two categories:

Edit: updated as of Philipp's comment below!

• Training

  • Vertex AI
  • [x] Custom Container Training Job to fine-tune (SFT) an LLM via TRL's CLI using QLoRA on a single GPU #44
  • [x] Custom Container Training Job to full fine-tune (SFT) an LLM via TRL's CLI on N GPUs (N > 1) via accelerate #44
  • [ ] Kubeflow Pipeline to fine-tune an LLM with a custom script on TPU (we already have https://github.com/huggingface/Google-Cloud-Containers/tree/main/examples/tpu-examples/causal-language-modeling)
  • GKE
  • [x] fine-tune (SFT) an LLM via TRL's CLI using QLoRA on a single GPU #53
  • [x] fine-tune (SFT) an LLM via TRL's CLI on N GPUs (N > 1) via accelerate #53

• Inference

  • Vertex AI (already a bunch of existing examples under examples/vertex-ai/notebooks only missing review / rewrite) #55
  • [x] PyTorch via transformers using a Custom Prediction Routine (CPR) in Vertex AI with any relevant model for any supported HF_TASK other than text-generation (could be either CPU, GPU or both)
  • [x] TGI via a pre-built DLC (on GPU) from the Hugging Face Hub
  • [x] TGI via a pre-built DLC (on GPU) from a GCS Bucket
  • [x] TEI via a pre-built DLC (could be either CPU, GPU or both)
  • [x] TEI via a pre-built DLC (could be either CPU, GPU or both) from a GCS Bucket
  • GKE
  • [x] TGI via a pre-built DLC with a custom Kubernetes configuration for GKE Autopilot mode (on GPU) from the Hugging Face Hub #41
  • [x] TGI via a pre-built DLC with a custom Kubernetes configuration for GKE Autopilot mode (on GPU) from a GCS Bucket #42
  • [x] TEI via a pre-built DLC with a custom Kubernetes configuration for GKE Autopilot mode (could be either CPU, GPU or both) #43
  • [x] TEI via a pre-built DLC with a custom Kubernetes configuration for GKE Autopilot mode (could be either CPU, GPU or both) from a GCS Bucket #42

[!NOTE] This issue assumes that the DLCs are already created and can be used as containers for the examples described above.

[philschmid]

Additionally, we should create examples for

• Training (GKE)

  • [x] fine-tune (SFT) an LLM via TRL's CLI using QLoRA on a single GPU
  • [x] fine-tune (SFT) an LLM via TRL's CLI using QLoRA on N GPUs (N > 1) via accelerate

• Inference

  • Vertex AI
  • [x] TGI via a pre-built DLC (on GPU) loading fine-tuned model from GCS
  • GKE
  • [x] TGI via a pre-built DLC (on GPU) loading fine-tuned model from GCS

[!NOTE] We already have a few examples covering some of those topics. Lets try to reuse most of them and rather update them then creating new versions.