Week 2

Finished

CWRU HPC Frequently Used Commands

Maintain a Session

Using Screen

NOTE: screen is available within the node where it is created. To login to hpcxx from another node, use ssh hpcxx.

  screen -S <session_name>    # Create a session named <session_name>
  Ctrl+A D                    # Detach from the screen but leave it running on the server
  screen -ls                  # List all sessions
  screen -r <session_name>    # Reattach to session <session_name>
  screen -r                   # Reattach, when there is only one session available

Using tmux
- tmux - Cheat Sheet

Start a Job (switch from login-node hpcxx to computer-node comptxx or GPU-node gputxx)
- To start a batch job with a .slurm script
```
  sbatch xxx.slurm
```
- To start an interactive job
```
  srun --pty /bin/bash
```
- To get email notification when the job start/end/fail/…
```
  # add these flags
  --mail-user=your_email@example.com
  --mail-type=ALL
```
Monitoring & Deleting Jobs [Job Control]
- For brief status of your jobs
```
  squeue -u zxc808
```
- For displaying job accounting information
```
  sacct -u zxc808
  sacct -u zxc808 -s pending # filter all the pending jobs
```
- For cancelling a specific job
```
  scancel -i <jobID>
```

Request GPU-nodes

To check the resource availability

To get information about GPU resources

  scontrol show node <gpu node>   # e.g. gput045

To request a GPU resource

  srun \
  -p gpu \        # change partition to gpu
  --gres=gpu:1 \  # request 1 GPU resource
  <other flags>
  --pty bash      # start an interactive job

Working with directory on CWRU HPC

NOTE: We are supposed to store our large files in our gallina home. However, gallina home is not accessible in compute/gpu nodes.
- Steps:
  - Create temporary scratch space
  - Copy data from RDS to /scratch
  - Run job script
  - Copy data from /scratch back to RDS
- Refer to Storing and using files on CWRU HPC and gallina
Use pip install in HPC

NOTE: Most of the time HPC will have the packages installed, we should be careful to avoid over-installing packages. If any package is not found, contact professor/mentor first, and then the support team.
- To check and load available versions of Python
```
  # check available versions
  module avail python/
  module spider python/

  # load specific version
  module load Python/3.11.3
```
- To install modules with pip
```
  pip install --user <package-name>
```

LlamaIndex + RAG

Studied Building Agentic RAG with LlamaIndex - DeepLearning.AI course
Learned about LLM Agents, read LLM Powered Autonomous Agents

Llama 2

Acquired License for Meta Llama 2 and 3
Attempted to run Llama 2 7B locally – failed due to lack of GPU

Other Messages From The Weekly Meeting

Priority: Focus on the generating side first, put off the evaluation side.
Finish something small but working first, add functionalities step by step.
Possible roadmap regarding generating frame blending examples:
- Run llama 2 with prompting
- Run llama 2 with prompting and RAG on FrameNet dataset
- Apply agents
- Fine-tuning llama 2 with FrameNet dataset
- … (maybe)

To-do List For Next Week

Read paper Lutma: A Frame-Making Tool for Collaborative FrameNet Development
Contact Arthur about running Llama 2 on HPC, since he might have done it so that I won’t have to do it all over again.
Try to run Llama 2 and generate some simple frame blending examples through prompts.

Weekly Meeting Slides Link

Red Hen Meeting - Jun 5

Week 1

Week 3