| # SPDX-License-Identifier: Apache-2.0 | |
| # Chakra 3 RIMAY — propose kernel | |
| # Leader: vllm-project/vllm @ 6548560 (Apache-2.0) | |
| # Distilled from: vllm/v1/sample/sampler.py Sampler.sample() | |
| import torch | |
| def propose(state, world, features, priors, seed: int) -> int: | |
| logits = torch.tensor(priors, dtype=torch.float32) | |
| logits = logits + torch.tensor(features, dtype=torch.float32) | |
| g = torch.Generator(); g.manual_seed(seed) | |
| probs = torch.softmax(logits, dim=-1) | |
| return int(torch.multinomial(probs, 1, generator=g).item()) | |