Debug: evidence-before-action investigation

The most expensive class of mistake in ML debugging is asserting a cause based on plausibility, then attempting a "fix" that masks the real problem. This skill enforces the discipline of probe → hypothesis → smoke → controls → claim, in that order.

The agentic Stop hook routes here from reason when an assistant claims a cause without backing tool output.

When to run

The user just said any of:

• "why is X failing / diverging / NaN / OOM / hung / slow / crashed"
• "the loss is going up", "metrics look weird", "GPU util is 0"
• "debug this", "diagnose", "troubleshoot", "investigate this run"
• pasted a log excerpt asking what's wrong

Five-step protocol

Step 1: cheap probes

Before forming any hypothesis, gather the cheap evidence. None of these cost more than a few seconds:

Process state:

ps aux | grep -E '(python|train|torchrun|accelerate)' | grep -v grep

Is the process still running? Zombie? Defunct? Multiple instances?

Kernel / system events:

dmesg | tail -100 # OOM kills, hardware errors, NFS errors
journalctl -xe --since "1 hour ago" | tail -50

GPU state:

nvidia-smi
nvidia-smi --query-gpu=utilization.gpu,memory.used,temperature.gpu --format=csv

Is the GPU even being used? Idle GPU during "training" means the process is blocked on data loading or has died.

Disk / filesystem:

df -h /path/to/run-dir
du -sh /path/to/run-dir/*

Out of disk? Checkpoints not being written?

Log scrollback: Read the last few hundred lines of the training log. Don't trust the user's summary, they may have skimmed. Look for:

• exception tracebacks
• repeated "loss=NaN" or "grad_norm=Inf"
• early-stop announcements (the run may have completed normally)
• the last successful epoch / step (where did progress stop)

Checkpoint state:

ls -la /path/to/run-dir/checkpoints/

When was the last checkpoint written? What does its size suggest? An empty .pt is different from a 2GB one cut short.

Step 2: hypothesis (labeled as hypothesis)

After the probe, state what might be happening, explicitly framed as a hypothesis:

"Hypothesis: the run is OOMing because dmesg shows oom-kill 3 minutes ago and the process is gone. Alternative hypotheses I haven't ruled out: (a) NFS write timeout, (b) explicit kill from a sibling process."

Never skip to "the cause is X." The hypothesis labels what you don't yet know.

Step 3: smoke run

The cheapest way to confirm or refute a hypothesis is to reproduce the failure shape under a controlled condition:

• OOM hypothesis: rerun with batch_size=1 for 1 step. If it survives, OOM is confirmed; if it fails the same way, OOM is wrong.
• Data hypothesis: rerun with a synthetic in-memory dataset. If it works, the data path is implicated.
• Model hypothesis: forward pass only on a single batch with eval() mode. Loss finite? Outputs sane?
• Optimizer hypothesis: rerun with lr=0. If the loss still explodes, the loss itself is broken (not the optimizer).
• Distributed hypothesis: rerun on 1 GPU. If it works, DDP / NCCL is implicated.

A 30-second smoke beats a 30-minute restart-and-pray.

Step 4: controls

If the smoke is ambiguous, run a control: change exactly one variable from the failing config and rerun the smoke. The differences narrow what mechanism is responsible.

Common control axes (change one at a time):

• single-source vs multi-source data
• default workers vs adjusted workers
• mixed-precision on vs off
• gradient checkpointing on vs off
• torch.compile on vs off

Step 5: claim cause

Only after evidence stacks up, probe, smoke, control, do you assert a cause. The claim should cite the specific tool output that proves it:

"Root cause: NFS write timeout. Evidence: dmesg shows nfs server X not responding at 14:23 (the same minute the last checkpoint was written), and the smoke with batch=1 reproduces the timeout. Recommended fix: bind-mount a local scratch dir for checkpoints and rsync to NFS at end of epoch."

If the evidence isn't stacking up, do not promote a hypothesis to a cause. Say "I don't yet know" and propose the next probe.

What to avoid

• "It's probably X, let me try Y" → no. Probe first.
• Restarting the run with a small change as the diagnostic. Smoke first, then restart deliberately.
• Citing only the user's narrative as evidence: re-read the actual log.
• Stopping at the first plausible cause when artifacts contradict it.

Output

A concise diagnostic report: (1) what the probes showed, (2) the hypothesis, (3) the smoke outcome, (4) the cause-or-uncertain verdict, (5) the recommended next action. Each claim cites the tool output that backs it.