Vector Addition 2^20 Throughput

Write a Triton implementation of element-wise vector addition for two contiguous float32 CUDA tensors of length 1048576.

Your submitted ZIP project must include solution.py at the project root. The file must define a Solution class with the Frontier-CS solve(spec_path) interface:

class Solution:
    def solve(self, spec_path: str | None = None) -> dict:
        return {"program_path": "kernel.py"}

The returned artifact must contain either {"code": "..."} or {"program_path": "..."}. The referenced module must define:

import torch
import triton
import triton.language as tl

def add(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    ...

add receives two CUDA tensors with the same shape and must return a CUDA tensor containing x + y. The evaluator calls Solution.solve(spec_path), materializes the returned artifact, and imports add inside a PyTorch/Triton environment created during the coexecuted-evaluator setup phase.

Scoring

Correctness is required. Incorrect output receives score 0.

The primary metric is score, a 0 to 100 throughput score normalized against CPU and PyTorch GPU baselines:

target = max(2 * pytorch_bandwidth_gbps / cpu_bandwidth_gbps, 1)
score = clamp(((custom_bandwidth_gbps / cpu_bandwidth_gbps - 1) / (target - 1)) * 100, 0, 100)

Constraints

• Public validation uses a tiny deterministic vector length. Official scoring uses the source 2^20 vector length.
• Inputs are contiguous CUDA float32 tensors.
• Correctness uses torch.allclose with rtol=1e-5 and atol=1e-8.
• Do not rely on network access or external services at evaluation time.
• The evaluator setup phase installs uv-managed CPython 3.12, PyTorch, and Triton with uv sync; your run-time code should only import packages available in that environment or files included in your ZIP project.

Coexecuted-Evaluator Boundary

This challenge uses coexecuted_benchmark, so the trusted coexecuted-evaluator imports participant code from /workspace inside the evaluator container. Official private benchmark configuration shares that container with participant code. The private asset contains no secrets.