| 1334 | self._init_samplers() |
| 1335 | |
| 1336 | def _init_samplers(self) -> None: |
| 1337 | for seq_id in range(self.n_seq_max): |
| 1338 | params = llama_cpp.llama_sampler_chain_default_params() |
| 1339 | params.no_perf = True |
| 1340 | sampler = llama_cpp.llama_sampler_chain_init(params) |
| 1341 | if self.top_k > 1: |
| 1342 | llama_cpp.llama_sampler_chain_add( |
| 1343 | sampler, |
| 1344 | llama_cpp.llama_sampler_init_top_k(self.top_k), |
| 1345 | ) |
| 1346 | llama_cpp.llama_sampler_chain_add( |
| 1347 | sampler, |
| 1348 | llama_cpp.llama_sampler_init_greedy(), |
| 1349 | ) |
| 1350 | self._samplers.append(sampler) |
| 1351 | |
| 1352 | def _passes_p_min(self, output_index: int) -> bool: |
| 1353 | if self.p_min <= 0.0: |