Validate num_inference_steps > 0 in DDPMScheduler.set_timesteps (#13394)

[jbbqqf]

Fixes #13394.

Summary

DDPMScheduler.set_timesteps(num_inference_steps=0) (or negative) silently accepted the input, ran np.linspace(0, num_train_timesteps - 1, 0) to produce an empty schedule, and left the scheduler in a broken state. The bug only surfaced later when step() operated on stale state — instead of failing fast.

`scheduling_block_refinement.py` already validates num_inference_steps > 0. This PR brings DDPM in line.

Fix

Add the same guard in DDPMScheduler.set_timesteps (and propagate to its # Copied from sibling `DDPMSchedulerParallel` via python utils/check_copies.py --fix_and_overwrite):

if num_inference_steps is None or num_inference_steps <= 0:
    raise ValueError(
        f"`num_inference_steps` must be a positive integer, got {num_inference_steps}."
    )

(The pre-existing too-large check is preserved.)

Reproduce BEFORE/AFTER yourself (copy-paste)

# --- BEFORE: origin/main, no error raised ---
git fetch origin main && git checkout origin/main
python - <<'PY'
from diffusers import DDPMScheduler
sched = DDPMScheduler()
sched.set_timesteps(num_inference_steps=0)
print(\"NO ERROR. timesteps:\", sched.timesteps)  # Expected: NO ERROR / empty
PY

# --- AFTER: this branch ---
git fetch origin pull/<PR>/head:fix && git checkout fix
python - <<'PY'
from diffusers import DDPMScheduler
sched = DDPMScheduler()
try:
    sched.set_timesteps(num_inference_steps=0)
except ValueError as e:
    print(f\"OK: {e}\")  # Expected: \`num_inference_steps\` must be a positive integer, got 0.
PY

What I ran locally

pytest tests/schedulers/test_scheduler_ddpm.py -k \"custom_timesteps or rejects_non_positive\"
# 5 passed

The new `test_set_timesteps_rejects_non_positive_num_inference_steps` covers both 0 and -1, and verifies the scheduler still works for a positive value afterwards. It fails on origin/main (no exception raised) and passes on this branch.

Notes

• 2 source files updated (DDPM + parallel sibling), 1 test file. Verified no other # Copied from consumers exist for DDPMScheduler.set_timesteps via python utils/check_copies.py.

---

🤖 Disclosure: Authored with assistance from Claude (Anthropic), reviewed and tested by me.