AReaL: Skill for Claude Code

Debug Distributed Training

Debugging guide for distributed training issues in AReaL (FSDP2, TP, CP, EP).

When to Use

This skill is triggered when:

• Training hangs or deadlocks
• Results differ across ranks or are numerically wrong
• OOM errors in distributed settings
• NCCL/communication errors or device mesh issues

Debugging Principles

Minimal Reproduction

Always follow the minimal demo principle: Reproduce with the least amount of code to narrow down the issue faster.

# Bad: Debug in full training loop
# Good: Create minimal script
import torch
import torch.distributed as dist

dist.init_process_group("nccl")
rank = dist.get_rank()

# Reproduce the exact operation that fails
tensor = torch.ones(10).cuda()
dist.all_reduce(tensor)  # <-- Isolate the failing op
print(f"Rank {rank}: {tensor}")

Reduction strategy:

1. Remove unrelated model components
2. Use small tensor sizes
3. Reduce world_size to minimum (e.g., 2 GPUs)
4. Remove torch.compile if possible
5. Disable activation checkpointing

Step-by-Step Debugging Guide

1. Hang Debugging (Deadlocks, Synchronization)

Environment Variables for Debugging:

# Full debug logging
export TORCH_DISTRIBUTED_DEBUG=DETAIL
export NCCL_DEBUG=INFO
export NCCL_DEBUG_SUBSYS=ALL

# torch.compile debugging
export TORCH_LOGS="+dynamo,recompiles"
export TORCHDYNAMO_VERBOSE=1

Dump Call Stack with py-spy (for hung processes):

# Find process IDs
ps aux | grep python

# Dump call stack of specific rank
py-spy dump --pid <PID>

# Record flame graph for performance analysis
py-spy record -o profile.svg --pid <PID> --duration 30

Common Causes:

1. Mismatched Collectives: One rank calls all_reduce, another doesn't.
2. Wrong Process Group: Using wrong group for collective.
3. Tensor Shape Mismatch: Different shapes across ranks.

Debug Steps:

# Verify group membership
mesh = parallel_dims.get_mesh("dp_shard_cp")
group = mesh.get_group()
print(f"Rank {dist.get_rank()}: group size = {dist.get_world_size(group)}")

# Print shapes on all ranks
print(f"Rank {dist.get_rank()}: tensor.shape = {tensor.shape}")
dist.barrier()