mirror of
https://github.com/allenai/olmocr.git
synced 2025-10-31 01:55:06 +00:00
350 lines
12 KiB
Bash
Executable File
350 lines
12 KiB
Bash
Executable File
#!/bin/bash
|
|
|
|
set -e
|
|
|
|
# Parse beaker-specific arguments
|
|
SKIP_DOCKER_BUILD=false
|
|
PREEMPTIBLE=false
|
|
EXP_NAME=""
|
|
NUM_GPUS=4
|
|
|
|
# Store all arguments to pass to python command
|
|
PYTHON_ARGS=()
|
|
|
|
while [[ $# -gt 0 ]]; do
|
|
case $1 in
|
|
--skip-docker-build)
|
|
SKIP_DOCKER_BUILD=true
|
|
shift
|
|
;;
|
|
--preemptible)
|
|
PREEMPTIBLE=true
|
|
shift
|
|
;;
|
|
--name)
|
|
EXP_NAME="$2"
|
|
shift 2
|
|
;;
|
|
--num-gpus)
|
|
NUM_GPUS="$2"
|
|
if [ "$NUM_GPUS" -lt 2 ] || [ "$NUM_GPUS" -gt 8 ]; then
|
|
echo "Error: --num-gpus must be between 2 and 8 (got: $NUM_GPUS)"
|
|
exit 1
|
|
fi
|
|
shift 2
|
|
;;
|
|
--help|-h)
|
|
echo "Usage: $0 [beaker-options] [grpo-training-options]"
|
|
echo ""
|
|
echo "Beaker-specific options:"
|
|
echo " --skip-docker-build Skip Docker build"
|
|
echo " --preemptible Use preemptible instances"
|
|
echo " --name NAME Experiment name (used in output directory)"
|
|
echo " --num-gpus N Number of GPUs to use (2-8, default: 4)"
|
|
echo ""
|
|
echo "All other arguments are forwarded to python -m olmocr.train.grpo_train"
|
|
echo "Run 'python -m olmocr.train.grpo_train --help' to see available training options"
|
|
echo ""
|
|
echo "This multi-GPU version runs:"
|
|
echo " - VLLM server on the last GPU"
|
|
echo " - Training on all other GPUs with DeepSpeed"
|
|
exit 0
|
|
;;
|
|
*)
|
|
# Store all other arguments to pass to python command
|
|
PYTHON_ARGS+=("$1")
|
|
shift
|
|
;;
|
|
esac
|
|
done
|
|
|
|
echo "Preemptible: $PREEMPTIBLE"
|
|
echo "Skip Docker Build: $SKIP_DOCKER_BUILD"
|
|
echo "Number of GPUs: $NUM_GPUS"
|
|
echo "Arguments to forward: ${PYTHON_ARGS[@]}"
|
|
|
|
# Use conda environment Python if available, otherwise use system Python
|
|
if [ -n "$CONDA_PREFIX" ]; then
|
|
PYTHON="$CONDA_PREFIX/bin/python"
|
|
echo "Using conda Python from: $CONDA_PREFIX"
|
|
else
|
|
PYTHON="python"
|
|
echo "Warning: No conda environment detected, using system Python"
|
|
fi
|
|
|
|
# Get version from version.py
|
|
VERSION=$($PYTHON -c 'import olmocr.version; print(olmocr.version.VERSION)')
|
|
echo "OlmOCR version: $VERSION"
|
|
|
|
# Get first 10 characters of git hash
|
|
GIT_HASH=$(git rev-parse HEAD | cut -c1-10)
|
|
echo "Git hash: $GIT_HASH"
|
|
|
|
# Get current git branch name
|
|
GIT_BRANCH=$(git rev-parse --abbrev-ref HEAD)
|
|
echo "Git branch: $GIT_BRANCH"
|
|
|
|
# Create full image tag
|
|
IMAGE_TAG="olmocr-grpo-${VERSION}-${GIT_HASH}"
|
|
echo "Building Docker image with tag: $IMAGE_TAG"
|
|
|
|
# Build and push Docker image if not skipping
|
|
if [ "$SKIP_DOCKER_BUILD" = false ]; then
|
|
echo "Building Docker image..."
|
|
docker build --platform linux/amd64 -f ./Dockerfile -t $IMAGE_TAG .
|
|
|
|
# Push image to beaker
|
|
echo "Trying to push image to Beaker..."
|
|
if ! beaker image create --workspace ai2/oe-data-pdf --name $IMAGE_TAG $IMAGE_TAG 2>/dev/null; then
|
|
echo "Warning: Beaker image with tag $IMAGE_TAG already exists. Using existing image."
|
|
fi
|
|
else
|
|
echo "Skipping Docker build as requested"
|
|
fi
|
|
|
|
# Get Beaker username
|
|
BEAKER_USER=$(beaker account whoami --format json | jq -r '.[0].name')
|
|
echo "Beaker user: $BEAKER_USER"
|
|
|
|
# Create Python script to run beaker experiment
|
|
cat << 'EOF' > /tmp/run_grpo_experiment_multi_gpu.py
|
|
import sys
|
|
import shlex
|
|
import os
|
|
from beaker import Beaker, ExperimentSpec, TaskSpec, TaskContext, ResultSpec, TaskResources, ImageSource, Priority, Constraints, EnvVar, DataMount
|
|
|
|
# Get parameters from command line
|
|
image_tag = sys.argv[1]
|
|
beaker_user = sys.argv[2]
|
|
git_branch = sys.argv[3]
|
|
git_hash = sys.argv[4]
|
|
preemptible = sys.argv[5] == "true"
|
|
exp_name = sys.argv[6] # Empty string if not provided
|
|
num_gpus = int(sys.argv[7])
|
|
# All remaining arguments are the python command arguments
|
|
python_args = sys.argv[8:]
|
|
|
|
# Calculate GPU assignments
|
|
vllm_gpu = num_gpus - 1 # Last GPU for VLLM
|
|
training_gpus = list(range(num_gpus - 1)) # All other GPUs for training
|
|
training_gpu_str = ",".join(str(g) for g in training_gpus)
|
|
num_training_processes = len(training_gpus)
|
|
|
|
# Initialize Beaker client
|
|
b = Beaker.from_env(default_workspace="ai2/olmocr")
|
|
|
|
# Process arguments to extract model path
|
|
model_sync_commands = []
|
|
modified_args = list(python_args)
|
|
model_path_local = None
|
|
for i in range(len(modified_args)):
|
|
if modified_args[i] == "--model_name" and i + 1 < len(modified_args):
|
|
model_path = modified_args[i + 1].rstrip('/')
|
|
if model_path.startswith("s3://"):
|
|
# Extract checkpoint name from S3 path (last part of path)
|
|
checkpoint_name = model_path.split('/')[-1]
|
|
local_model_path = f"/data/models/{checkpoint_name}"
|
|
model_path_local = local_model_path
|
|
|
|
# Create sync commands
|
|
model_sync_commands = [
|
|
f"echo 'Syncing model from S3: {model_path}'",
|
|
"mkdir -p /data/models",
|
|
f"s5cmd sync '{model_path}/*' '{local_model_path}/'",
|
|
]
|
|
|
|
# Replace S3 path with local path in arguments
|
|
modified_args[i + 1] = local_model_path
|
|
else:
|
|
model_path_local = model_path
|
|
break
|
|
|
|
# Build setup commands
|
|
setup_commands = [
|
|
# Install dependencies
|
|
"pip install .[train]",
|
|
"pip install trl==0.23.0 wandb",
|
|
"pip install transformers==4.55.2", # Updated for GRPO compatibility
|
|
"pip install flash-attn==2.8.0.post2 --no-build-isolation",
|
|
"pip install vllm==v0.10.1.1",
|
|
"pip install s5cmd",
|
|
"pip install accelerate deepspeed",
|
|
|
|
# Sync the bench data from S3
|
|
"echo 'Syncing bench data from S3...'",
|
|
"mkdir -p /data/olmOCR-bench",
|
|
"s5cmd sync 's3://ai2-oe-data/jakep/olmocr/olmOCR-bench-snapshot-082225/*' /data/olmOCR-bench/",
|
|
"s5cmd sync 's3://ai2-oe-data/jakep/grpo_data_mixes/*' /data/jakep/grpo_data_mixes/",
|
|
]
|
|
|
|
# Add model sync commands if needed
|
|
if model_sync_commands:
|
|
setup_commands.extend(model_sync_commands)
|
|
|
|
# Determine model path for VLLM server
|
|
if model_path_local:
|
|
vllm_model_arg = model_path_local
|
|
else:
|
|
# Default model if not specified
|
|
vllm_model_arg = "Qwen/Qwen2.5-VL-7B-Instruct"
|
|
for i, arg in enumerate(modified_args):
|
|
if arg == "--model_name" and i + 1 < len(modified_args):
|
|
vllm_model_arg = modified_args[i + 1]
|
|
break
|
|
|
|
# Extract gradient_accumulation_steps from arguments if provided, otherwise use default
|
|
grad_acc_steps = 8 # Default value
|
|
for i, arg in enumerate(modified_args):
|
|
if arg == "--gradient_accumulation_steps" and i + 1 < len(modified_args):
|
|
try:
|
|
grad_acc_steps = int(modified_args[i + 1])
|
|
except (ValueError, IndexError):
|
|
pass # Keep default if parsing fails
|
|
break
|
|
|
|
# Build the GRPO training command with forwarded arguments
|
|
# Force --vllm_mode server
|
|
grpo_cmd = f"CUDA_VISIBLE_DEVICES={training_gpu_str} accelerate launch --use_deepspeed --zero_stage 2 --num_processes {num_training_processes} --gradient_accumulation_steps {grad_acc_steps} -m olmocr.train.grpo_train"
|
|
|
|
# Add --vllm_mode server if not already in arguments
|
|
arg_str = " ".join(modified_args)
|
|
if "--vllm_mode" not in arg_str:
|
|
grpo_cmd += " --vllm_mode server"
|
|
|
|
# Check if certain required arguments are in the provided args, add defaults if not
|
|
if "--train_bench_data_folder" not in arg_str:
|
|
grpo_cmd += " --train_bench_data_folder /data/olmOCR-bench/bench_data"
|
|
if "--eval_bench_data_folder" not in arg_str:
|
|
grpo_cmd += " --eval_bench_data_folder /data/olmOCR-bench/bench_data"
|
|
if "--output_dir" not in arg_str:
|
|
output_dir = "/weka/oe-training-default/jakep/olmocr-grpo-checkpoints"
|
|
# Build subdirectory based on exp_name and BEAKER_WORKLOAD_ID
|
|
beaker_workload_id = "${BEAKER_WORKLOAD_ID}"
|
|
if exp_name:
|
|
# For multi-GPU runs, add suffix to distinguish
|
|
output_dir = f"{output_dir}/{exp_name}-multigpu-{beaker_workload_id}"
|
|
else:
|
|
output_dir = f"{output_dir}/multigpu-{beaker_workload_id}"
|
|
grpo_cmd += f" --output_dir {output_dir}"
|
|
|
|
# Add all the (possibly modified) arguments, filtering out --vllm_mode if it exists to avoid duplicates
|
|
# Note: We keep --gradient_accumulation_steps in the args even though we use it for accelerate,
|
|
# because the training script also needs it for its configuration
|
|
filtered_args = []
|
|
skip_next = False
|
|
for i, arg in enumerate(modified_args):
|
|
if skip_next:
|
|
skip_next = False
|
|
continue
|
|
if arg == "--vllm_mode":
|
|
skip_next = True # Skip this and the next argument
|
|
continue
|
|
filtered_args.append(arg)
|
|
|
|
grpo_cmd += " " + " ".join(filtered_args)
|
|
|
|
# Create a bash script as a single command string
|
|
bash_script = f"""
|
|
set -e
|
|
|
|
# Setup commands
|
|
{" && ".join(setup_commands)}
|
|
|
|
# Start VLLM server in background (output goes to console)
|
|
echo 'Starting VLLM server on GPU {vllm_gpu} as background process...'
|
|
CUDA_VISIBLE_DEVICES={vllm_gpu} trl vllm-serve --model {vllm_model_arg} --port 8000 --gpu-memory-utilization 0.5 &
|
|
VLLM_PID=$!
|
|
echo "VLLM server started with PID: $VLLM_PID"
|
|
|
|
# Wait for VLLM server to be ready
|
|
echo 'Waiting for VLLM server to be ready...'
|
|
sleep 30
|
|
for i in {{1..60}}; do
|
|
if curl -s http://localhost:8000/health; then
|
|
echo ' - VLLM server is ready!'
|
|
break
|
|
else
|
|
echo 'Still waiting for VLLM server...'
|
|
sleep 5
|
|
fi
|
|
done
|
|
|
|
# Run training
|
|
echo 'Starting GRPO training on GPUs {training_gpu_str}...'
|
|
{grpo_cmd}
|
|
|
|
# Cleanup
|
|
echo 'Training completed. Killing VLLM server...'
|
|
kill $VLLM_PID || true
|
|
echo 'VLLM server stopped.'
|
|
"""
|
|
|
|
# Create single task spec
|
|
task_spec = TaskSpec(
|
|
name="olmocr-grpo-multi-gpu",
|
|
image=ImageSource(beaker=f"{beaker_user}/{image_tag}"),
|
|
command=[
|
|
"bash", "-c",
|
|
bash_script
|
|
],
|
|
context=TaskContext(
|
|
priority=Priority.normal,
|
|
preemptible=preemptible,
|
|
),
|
|
resources=TaskResources(
|
|
gpu_count=num_gpus, # Request the specified number of GPUs
|
|
shared_memory="10GiB"
|
|
),
|
|
constraints=Constraints(cluster=["ai2/jupiter", "ai2/saturn"]),
|
|
result=ResultSpec(path="/noop-results"),
|
|
env_vars=[
|
|
EnvVar(name="LOG_FILTER_TYPE", value="local_rank0_only"),
|
|
EnvVar(name="OMP_NUM_THREADS", value="8"),
|
|
EnvVar(name="BEAKER_USER_ID", value=beaker_user),
|
|
EnvVar(name="AWS_ACCESS_KEY_ID", secret="ALLENNLP_AWS_ACCESS_KEY_ID"),
|
|
EnvVar(name="AWS_SECRET_ACCESS_KEY", secret="ALLENNLP_AWS_SECRET_ACCESS_KEY"),
|
|
EnvVar(name="WANDB_API_KEY", secret="JAKE_WANDB_API_KEY"),
|
|
],
|
|
datasets=[
|
|
DataMount.new(mount_path="/weka/oe-data-default", weka="oe-data-default"),
|
|
DataMount.new(mount_path="/weka/oe-training-default", weka="oe-training-default"),
|
|
]
|
|
)
|
|
|
|
# Extract model name from arguments if provided (for description)
|
|
model_name = "Unknown"
|
|
for i, arg in enumerate(modified_args):
|
|
if arg in ["--model_name", "--model"]:
|
|
if i + 1 < len(modified_args):
|
|
model_name = modified_args[i + 1]
|
|
break
|
|
|
|
# Create experiment spec with single task
|
|
experiment_spec = ExperimentSpec(
|
|
description=f"OlmOCR GRPO Multi-GPU Training ({num_training_processes} GPUs + VLLM Server) - Model: {model_name}, Branch: {git_branch}, Commit: {git_hash}",
|
|
budget="ai2/oe-base",
|
|
tasks=[task_spec], # Single task that manages both VLLM and training
|
|
)
|
|
|
|
# Create the experiment
|
|
experiment = b.experiment.create(spec=experiment_spec, workspace="ai2/olmocr")
|
|
print(f"Created multi-GPU GRPO training experiment: {experiment.id}")
|
|
print(f"View at: https://beaker.org/ex/{experiment.id}")
|
|
EOF
|
|
|
|
# Run the Python script to create the experiment
|
|
echo "Creating Beaker multi-GPU GRPO experiment..."
|
|
$PYTHON /tmp/run_grpo_experiment_multi_gpu.py \
|
|
"$IMAGE_TAG" \
|
|
"$BEAKER_USER" \
|
|
"$GIT_BRANCH" \
|
|
"$GIT_HASH" \
|
|
"$PREEMPTIBLE" \
|
|
"$EXP_NAME" \
|
|
"$NUM_GPUS" \
|
|
"${PYTHON_ARGS[@]}"
|
|
|
|
# Clean up temporary file
|
|
rm /tmp/run_grpo_experiment_multi_gpu.py
|
|
|
|
echo "Multi-GPU GRPO training experiment submitted successfully!" |