Add multi-GPU vector addition challenge + script support

[kunal-mansukhani]

Summary

New Pro-only multi-GPU test challenge at challenges/easy/100_multi_gpu_vector_add (num_gpus=2), with starter + solution for all 5 accelerated languages (PyTorch, CUDA, Triton, CuTe, JAX). Used to validate the multi-GPU runner path end-to-end in the companion infra PR (AlphaGPU/leetgpu-infra#333).

Plus two script updates that were needed to make multi-GPU submissions work via run_challenge.py.

Changes

New challenge: 100_multi_gpu_vector_add

• challenge.py: num_gpus=2, access_tier="pro". Reference implementation is single-device (torch.add(A, B, out=C)) — the runner runs it independently on each rank and validates per-rank outputs.
• challenge.html: spec explaining the fully-replicated input/output model and the env vars exposed to solve() (RANK, WORLD_SIZE, LOCAL_RANK, LEETGPU_NCCL_ID_FILE).
• Starters + solutions for PyTorch (dist.all_reduce), CUDA (NCCL via LEETGPU_NCCL_ID_FILE), Triton (@triton.jit slice kernel + dist.all_reduce), CuTe (plain Python host + dist.all_reduce), JAX (trivial, validates jax.distributed.initialize).

scripts/update_challenges.py

• Forward num_gpus from ChallengeBase to the backend (field was declared but never propagated).

scripts/run_challenge.py

• Auto-detect num_gpus from challenge.py via regex and send gpuCount in the submission payload; new --gpu-count override flag.
• Prefer language-tagged solution filenames (solution.triton.py, solution.cute.py, solution.jax.py) so multiple Python-based languages can coexist in one solution/ dir. Falls back to solution.<ext>.
• Print all stdout + stderr WebSocket frames (was silently dropping type=stderr output).
• Terminate on test-case-failed / compilation-failed / tampering-detected / out-of-memory in addition to the previous terminal statuses.

Test plan

All validated end-to-end against the companion infra branch on real Modal hardware:

• PyTorch — run + full submit on 2× T4 and 4× T4
• CUDA — run + full submit on 2× T4
• Triton — run + full submit on 2× T4
• CuTe — run + full submit on 2× T4
• JAX — run + full submit on 2× T4

Notes

• The JAX solution is intentionally trivial (each rank computes the full result locally). It validates that jax.distributed.initialize succeeds and the output flows back correctly; a real distributed JAX pattern would use jax.experimental.multihost_utils / pjit.
• This PR depends on AlphaGPU/leetgpu-infra#333 — without the infra changes, the script will try to submit with gpuCount=2 and the server will reject/coerce it.

🤖 Generated with Claude Code