Sagemaker integration (#151)

* remove relative imports

* sm-integration

* folder rename + args for region, arn, s3

* setup.py fixes

* cleanup tri mentions + address nits

* Updates for sagemaker integration

* sm fixes

* format + conflicts

---------

Co-authored-by: Achal Dave <[email protected]>
Co-authored-by: Achal Dave <[email protected]>

.dockerignore

@@ -0,0 +1,4 @@
+venv
+wandb
+logs
+checkpoints

.gitignore

@@ -139,5 +139,6 @@ weights*
 out*
 tests/assets/*
 .vscode/
+secrets.env
 checkpoints/
 experiments/

sagemaker_train/Dockerfile

@@ -0,0 +1,34 @@
+ARG AWS_REGION
+
+# SageMaker PyTorch image
+FROM 763104351884.dkr.ecr.${AWS_REGION}.amazonaws.com/pytorch-training:2.1.0-gpu-py310-cu121-ubuntu20.04-sagemaker
+
+# Run custom installation of libraries
+# RUN pip install xxx
+# RUN apt-get update && apt-get install -y xxx
+# ENV <your environment variables>
+# etc....
+
+# Remove the conda installed symlink for libcurl, which causes an error with curl.
+# Fixes the following error:
+# curl: /opt/conda/lib/libcurl.so.4: no version information available (required by curl)
+RUN rm /opt/conda/lib/libcurl.so.4
+
+ENV PATH="/opt/ml/code:${PATH}"
+
+# this environment variable is used by the SageMaker PyTorch container to determine our user code directory.
+ENV SAGEMAKER_SUBMIT_DIRECTORY /opt/ml/code
+
+# /opt/ml and all subdirectories are utilized by SageMaker, use the /code subdirectory to store your user code.
+COPY . /opt/ml/code/
+RUN rm /opt/ml/code/setup.py
+
+RUN pip install -r /opt/ml/code/requirements.txt
+RUN pip uninstall flash-attn -y
+RUN pip install flash-attn>=2.2
+# # Prevent sagemaker from installing requirements again.
+# RUN rm /opt/ml/code/setup.py
+RUN rm /opt/ml/code/requirements.txt
+
+# Defines a script entrypoint 
+ENV SAGEMAKER_PROGRAM open_lm/main.py

sagemaker_train/Dockerfile_update

@@ -0,0 +1,15 @@
+ARG BASE_DOCKER
+# Dockerfile that updates the container with new code.
+# SageMaker PyTorch image
+FROM ${BASE_DOCKER}
+
+# /opt/ml and all subdirectories are utilized by SageMaker, use the /code subdirectory to store your user code.
+COPY . /opt/ml/code/
+
+# RUN pip install -e /opt/ml/code/
+
+# Prevent sagemaker from installing requirements again.
+RUN rm /opt/ml/code/setup.py
+RUN rm /opt/ml/code/requirements.txt
+
+ENV SAGEMAKER_PROGRAM open_lm/main.py

sagemaker_train/cfg_sample.yaml

@@ -0,0 +1,38 @@
+accum-freq: 4
+beta1: 0.9
+beta2: 0.95
+data-key: "json"
+dataset-resampled: True
+# delete-previous-checkpoint: False
+# Total 25B * 40 = 1T tokens
+epochs: 40
+fsdp: True
+fsdp-limit-all-gathers: True
+# grad-checkpointing: False
+grad-clip-norm: 1
+log-every-n-steps: 20
+model: "open_lm_7b"
+name: "sample_7b"
+precision: "amp_bfloat16"
+report-to: "wandb"
+seed: 124
+train-data-mix-weights: [0.725, 0.275]
+train-data: ["TODO"]
+train-num-samples: 25_000_000_000
+wandb-project-name: "lm1"
+workers: 4
+logs: /opt/ml/checkpoints/
+
+# Some important parameters, double checked with Mitchell:
+batch-size: 16
+ffn-type: swiglu
+# fsdp-amp: False
+fsdp-pure-bf16: True
+fsdp-backward-prefetch: True
+lr: 3.e-4
+lr-cooldown-end: 3.e-5
+model-norm: "gain_only_lp_layer_norm"
+qk-norm: True
+warmup: 5000
+wd: 0.1
+z-loss-coefficient: 1.e-4

sagemaker_train/launch_sagemaker_train.py

@@ -0,0 +1,194 @@
+import argparse
+import time
+import os
+import subprocess
+from datetime import datetime
+from pathlib import Path
+
+import boto3
+import sagemaker
+from sagemaker.pytorch import PyTorch
+
+
+NAME = "openlm-main"
+INSTANCE_MAPPER = {
+    "p4": "ml.p4d.24xlarge",
+    "p4de": "ml.p4de.24xlarge",
+    "p5": "ml.p5.48xlarge",
+}
+
+
+def run_command(command):
+    print(f"=> {command}")
+    subprocess.run(command, shell=True, check=True)
+
+
+def get_image(user, instance_type, build_type=None, profile="poweruser", region="us-east-1"):
+    os.environ["AWS_PROFILE"] = f"{profile}"
+    account = subprocess.getoutput(
+        f"aws --region {region} --profile {profile} sts get-caller-identity --query Account --output text"
+    )
+    docker_dir = Path(__file__).parent
+    if instance_type in ("p4", "p4de"):
+        algorithm_name = f"{user}-{NAME}-p4"
+        dockerfile_base = docker_dir / "Dockerfile"
+        dockerfile_update = docker_dir / "Dockerfile_update"
+    elif instance_type == "p5":
+        algorithm_name = f"{user}-{NAME}-p5"
+        dockerfile_base = docker_dir / "Dockerfile"
+        dockerfile_update = docker_dir / "Dockerfile_update"
+    else:
+        raise ValueError(f"Unknown instance_type: {instance_type}")
+    fullname = f"{account}.dkr.ecr.{region}.amazonaws.com/{algorithm_name}:latest"
+    if build_type is None:
+        return fullname
+
+    login_cmd = f"aws ecr get-login-password --region {region} --profile {profile} | docker login --username AWS --password-stdin"
+
+    if build_type == "full":
+        print("Building container")
+        commands = [
+            # Log in to Sagemaker account to get image.
+            f"{login_cmd} 763104351884.dkr.ecr.{region}.amazonaws.com",
+            f"docker build --progress=plain -f {dockerfile_base} --build-arg AWS_REGION={region} -t {algorithm_name} .",
+            f"docker tag {algorithm_name} {fullname}",
+            f"{login_cmd} {fullname}",
+            (
+                f"aws --region {region} ecr describe-repositories --repository-names {algorithm_name} || "
+                f"aws --region {region} ecr create-repository --repository-name {algorithm_name}"
+            ),
+        ]
+    elif build_type == "update":
+        print("Updating container")
+        commands = [
+            f"docker build --progress=plain -f {dockerfile_update} --build-arg BASE_DOCKER={algorithm_name} -t {algorithm_name} .",
+            f"docker tag {algorithm_name} {fullname}",
+            f"{login_cmd} {fullname}",
+        ]
+    else:
+        raise ValueError(f"Unknown build_type: {build_type}")
+
+    # Create command, making sure to exit if any part breaks.
+    command = "\n".join([f"{x} || exit 1" for x in commands])
+    run_command(command)
+    run_command(f"docker push {fullname}")
+    print("Sleeping for 5 seconds to ensure push succeeded")
+    time.sleep(5)
+    return f"{account}.dkr.ecr.{region}.amazonaws.com/{algorithm_name}:latest"
+
+
+def main():
+    # Use first line of file docstring as description if it exists.
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--build-type", choices=["full", "update"], help="Build image from scratch")
+    parser.add_argument("--local", action="store_true")
+    parser.add_argument("--user", required=True, help="User name")
+    parser.add_argument("--cfg-path", required=True, help="Launch config")
+
+    # AWS profile args
+    parser.add_argument("--region", default="us-east-1", help="AWS region")
+    parser.add_argument("--profile", default="poweruser", help="AWS profile to use")
+    parser.add_argument("--arn", default=None, help="If None, reads from SAGEMAKER_ARN env var")
+    parser.add_argument(
+        "--s3-remote-sync", default=None, help="S3 path to sync to. If none, reads from S3_REMOTE_SYNC env var"
+    )
+
+    # Instance args
+    parser.add_argument("--instance-count", default=1, type=int, help="Number of instances")
+    parser.add_argument("--instance-type", default="p4de", choices=list(INSTANCE_MAPPER.keys()))
+    parser.add_argument("--spot-instance", action="store_true")
+
+    args = parser.parse_args()
+    main_after_setup_move(args)
+
+
+def main_after_setup_move(args):
+    if args.arn is None:
+        assert "SAGEMAKER_ARN" in os.environ, "Please specify --arn or set the SAGEMAKER_ARN environment variable"
+        args.arn = os.environ["SAGEMAKER_ARN"]
+
+    if args.s3_remote_sync is None:
+        assert (
+            "S3_REMOTE_SYNC" in os.environ
+        ), "Please specify --s3-remote-sync or set the S3_REMOTE_SYNC environment variable"
+        args.s3_remote_sync = os.environ["S3_REMOTE_SYNC"]
+
+    image = get_image(
+        args.user,
+        args.instance_type,
+        region=args.region,
+        build_type=args.build_type,
+        profile=args.profile,
+    )
+
+    ##########
+    # Create session and make sure of account and region
+    ##########
+    sagemaker_session = sagemaker.Session(boto_session=boto3.session.Session(region_name=args.region))
+
+    role = args.arn
+    # provide a pre-existing role ARN as an alternative to creating a new role
+    role_name = role.split(["/"][-1])
+    print(f"SageMaker Execution Role:{role}")
+    print(f"The name of the Execution role: {role_name[-1]}")
+
+    client = boto3.client("sts")
+    account = client.get_caller_identity()["Account"]
+    print(f"AWS account:{account}")
+
+    session = boto3.session.Session()
+    region = session.region_name
+    print(f"AWS region:{region}")
+
+    ##########
+    # Configure the training
+    ##########
+    base_job_name = f"{args.user.replace('.', '-')}-{NAME}"
+
+    checkpoint_local_path = "/opt/ml/checkpoints"
+
+    def get_job_name(base):
+        now = datetime.now()
+        # Format example: 2023-03-03-10-14-02-324
+        now_ms_str = f"{now.microsecond // 1000:03d}"
+        date_str = f"{now.strftime('%Y-%m-%d-%H-%M-%S')}-{now_ms_str}"
+
+        job_name = "_".join([base, date_str])
+
+        return job_name
+
+    job_name = get_job_name(base_job_name)
+
+    output_root = f"{args.s3_remote_sync}/sagemaker/{args.user}/{NAME}/"
+    output_s3 = os.path.join(output_root, job_name)
+
+    estimator = PyTorch(
+        entry_point="open_lm/main.py",
+        sagemaker_session=sagemaker_session,
+        base_job_name=base_job_name,
+        hyperparameters={"config": args.cfg_path},
+        role=role,
+        image_uri=image,
+        instance_count=args.instance_count,
+        instance_type="local_gpu" if args.local else INSTANCE_MAPPER[args.instance_type],
+        train_use_spot_instances=args.spot_instance,
+        output_path=output_s3,
+        job_name=job_name,
+        checkpoint_s3_uri=None if args.local else f"{output_s3}/checkpoint",
+        checkpoint_local_path=None if args.local else checkpoint_local_path,
+        code_location=output_s3,
+        # Training using SMDataParallel Distributed Training Framework
+        distribution={"torch_distributed": {"enabled": True}},
+        # Max run 5 days
+        max_run=5 * 24 * 60 * 60,
+        max_wait=5 * 24 * 60 * 60 if args.spot_instance else None,
+        input_mode="FastFile",
+        # environment={"TORCH_DISTRIBUTED_DEBUG": "DETAIL", "TORCH_CPP_LOG_LEVEL": "INFO"},
+        keep_alive_period_in_seconds=30 * 60 if not args.spot_instance else None,  # 30 minutes
+    )
+
+    estimator.fit()
+
+
+if __name__ == "__main__":
+    main()