Fix/causal conv1d cuda128

[dzorlu]

---

[gemini-code-assist]

Code Review

This pull request introduces a massive restructuring of the project, moving towards a modular skyrl full-stack library and adding the skyrl-agent component. It includes significant updates to documentation, Docker infrastructure, and adds numerous examples and data processing scripts. The changes are extensive and set a new foundation for the project. My review focuses on improving maintainability and portability by addressing issues like code duplication in Dockerfiles and hardcoded paths in scripts and documentation. I've also identified a potential bug in the GHA runner health check script.

[gemini-code-assist]

There is significant code duplication between this Dockerfile and docker/Dockerfile. The first 20 lines are nearly identical. This makes maintenance difficult, as changes need to be applied in multiple places. Consider creating a common base Dockerfile (e.g., Dockerfile.base) and have Dockerfile, Dockerfile.megatron, and Dockerfile.ray244 build FROM it. This would centralize the common setup steps like installing system packages, CUDA toolkit, and uv.

[gemini-code-assist]

This Dockerfile duplicates a large portion of setup code from docker/Dockerfile and docker/Dockerfile.megatron. To improve maintainability and reduce redundancy, it's highly recommended to use a multi-stage build or a common base image that contains all the shared setup instructions.

[gemini-code-assist]

The log function attempts to write to /var/log/runner-health-check.log using sudo tee. However, the setup-gha-runner.sh script installs the cron job for the current user, not root. When the cron job runs, the sudo command will fail because it cannot ask for a password, leading to the health check script failing to log anything. The log file should be written to a user-writable location, or the cron job should be installed in the system-wide crontab to run as root.

Suggested change

	log() {
	echo "[$(date '+%Y-%m-%d %H:%M:%S')] $1" | sudo tee -a "$LOG_FILE" >/dev/null
	}
	log() {
	echo "[$(date '+%Y-%m-%d %H:%M:%S')] $1" >> "$LOG_FILE"
	}

[gemini-code-assist]

The file paths for fleet-sdk and OpenEnv are hardcoded to a specific user's local directory (/Users/deniz/...). This makes the instructions not generally usable for other developers. These should be replaced with placeholder paths or instructions on how to set them up.

Suggested change

	- Fleet SDK repo: `/Users/deniz/repos/fleet-sdk`
	- OpenEnv repo: `/Users/deniz/repos/OpenEnv`
	- Fleet SDK repo: `<path/to/your/local/fleet-sdk>`
	- OpenEnv repo: `<path/to/your/local/OpenEnv>`

[gemini-code-assist]

The documentation contains hardcoded AWS resource IDs, such as AMI ID, security group ID, subnet ID, and instance IDs (also on line 76). This makes the commands not directly usable and prone to copy-paste errors. These should be replaced with placeholders (e.g., <YOUR_AMI_ID>) to make it clear that users need to substitute their own values.

Suggested change

	aws ec2 run-instances --image-id ami-0c7217cdde317cfec --instance-type t3.xlarge --key-name gha-runner-key --security-group-ids sg-00fefd8181d51909d --subnet-id subnet-03879810067f57f85 --block-device-mappings '[{"DeviceName":"/dev/sda1","Ebs":{"VolumeSize":50,"VolumeType":"gp3"}}]' --tag-specifications 'ResourceType=instance,Tags=[{Key=Name,Value=fleet-runner-N}]'
	aws ec2 run-instances --image-id <YOUR_AMI_ID> --instance-type t3.xlarge --key-name <YOUR_KEY_NAME> --security-group-ids <YOUR_SECURITY_GROUP_ID> --subnet-id <YOUR_SUBNET_ID> --block-device-mappings '[{"DeviceName":"/dev/sda1","Ebs":{"VolumeSize":50,"VolumeType":"gp3"}}]' --tag-specifications 'ResourceType=instance,Tags=[{Key=Name,Value=fleet-runner-N}]'

[gemini-code-assist]

The path to the dataset is hardcoded. This makes the script difficult to run in different environments. This path should be configurable, for example, via a command-line argument or an environment variable.

Suggested change

	dataset_file = "/data/sycao/r2e-all/train.parquet"
	dataset_file = os.environ.get("DATASET_FILE", "/data/sycao/r2e-all/train.parquet")

[gemini-code-assist]

The script contains hardcoded paths for datasets and checkpoint directories (e.g., /mnt/shared_storage/datasets/r2e-all, $HOME/ckpts). This makes the script not portable and difficult for other developers to run without modification. It's better to use environment variables for these paths with sensible defaults, allowing users to easily override them.

Suggested change

	DATA_DIR="/mnt/shared_storage/datasets/r2e-all"
	TRAIN_DATA="${DATA_DIR}/train.parquet"
	VAL_DATA="${DATA_DIR}/validation.parquet"
	CKPT_DIR=$HOME/ckpts
	EXPORT_DIR=$HOME/exports
	DATA_DIR="${DATA_DIR:-/mnt/shared_storage/datasets/r2e-all}"
	TRAIN_DATA="${TRAIN_DATA:-${DATA_DIR}/train.parquet}"
	VAL_DATA="${VAL_DATA:-${DATA_DIR}/validation.parquet}"
	CKPT_DIR="${CKPT_DIR:-$HOME/ckpts}"
	EXPORT_DIR="${EXPORT_DIR:-$HOME/exports}"

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.