| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Shared BF16 baselines for the quantization comparison tests. |
| 4 | |
| 5 | Several tests here quantize the same model and compare against an |
| 6 | unquantized run of it. Computing that reference once per session avoids |
| 7 | booting an identical engine for every quantization scheme under test. |
| 8 | |
| 9 | Only the outputs are cached; each engine is still torn down as usual, so |
| 10 | tests keep the isolation ``VllmRunner.__exit__`` provides. |
| 11 | """ |
| 12 | |
| 13 | import pytest |
| 14 | |
| 15 | |
| 16 | @pytest.fixture(scope="session") |
| 17 | def quant_baseline_logprobs(vllm_runner): |
| 18 | """Greedy logprobs for an unquantized model, computed once per config.""" |
| 19 | cache: dict = {} |
| 20 | |
| 21 | def baseline( |
| 22 | model: str, |
| 23 | prompts: list[str], |
| 24 | *, |
| 25 | max_model_len: int, |
| 26 | max_tokens: int, |
| 27 | num_logprobs: int, |
| 28 | ): |
| 29 | key = (model, max_model_len, max_tokens, num_logprobs, tuple(prompts)) |
| 30 | if key not in cache: |
| 31 | with vllm_runner( |
| 32 | model, |
| 33 | max_model_len=max_model_len, |
| 34 | enforce_eager=True, |
| 35 | ) as vllm_model: |
| 36 | cache[key] = vllm_model.generate_greedy_logprobs( |
| 37 | prompts, max_tokens, num_logprobs |
| 38 | ) |
| 39 | return cache[key] |
| 40 | |
| 41 | return baseline |