|
| 1 | +# SPDX-License-Identifier: Apache-2.0 |
| 2 | +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | + |
| 4 | +import pytest |
| 5 | + |
| 6 | +from vllm import SamplingParams |
| 7 | +from vllm.logprobs import FlattenLogprobs |
| 8 | + |
| 9 | +MODELS = ["distilbert/distilgpt2"] |
| 10 | +MAX_TOKENS = 5 |
| 11 | +NUM_TOP_LOGPROBS = 5 |
| 12 | +NUM_PROMPT_LOGPROBS = 7 |
| 13 | +MAX_LOGPROBS = max(NUM_TOP_LOGPROBS, NUM_PROMPT_LOGPROBS) |
| 14 | + |
| 15 | + |
| 16 | +@pytest.mark.parametrize("model", MODELS) |
| 17 | +@pytest.mark.parametrize("dtype", ["half"]) |
| 18 | +@pytest.mark.parametrize("greedy", [True, False]) |
| 19 | +@pytest.mark.parametrize("flatten_logprobs", [True, False]) |
| 20 | +def test_ranks( |
| 21 | + vllm_runner, |
| 22 | + model, |
| 23 | + dtype, |
| 24 | + greedy, |
| 25 | + flatten_logprobs, |
| 26 | + example_prompts, |
| 27 | + monkeypatch: pytest.MonkeyPatch, |
| 28 | +): |
| 29 | + monkeypatch.setenv("VLLM_FLATTEN_LOGPROBS", "1" if flatten_logprobs else "0") |
| 30 | + with vllm_runner(model, dtype=dtype, max_logprobs=MAX_LOGPROBS) as vllm_model: |
| 31 | + tokenizer = vllm_model.llm.get_tokenizer() |
| 32 | + example_prompt_tokens = [tokenizer.encode(prompt) for prompt in example_prompts] |
| 33 | + sampling_params = SamplingParams( |
| 34 | + temperature=0.0 if greedy else 1.0, |
| 35 | + top_p=1.0, |
| 36 | + max_tokens=MAX_TOKENS, |
| 37 | + logprobs=NUM_TOP_LOGPROBS, |
| 38 | + prompt_logprobs=NUM_PROMPT_LOGPROBS, |
| 39 | + ) |
| 40 | + results = vllm_model.generate_w_logprobs(example_prompts, sampling_params) |
| 41 | + |
| 42 | + assert len(results) == len(example_prompt_tokens) |
| 43 | + for i, (result, prompt_tokens) in enumerate(zip(results, example_prompt_tokens)): |
| 44 | + decode_tokens, _, decode_logprobs, prompt_logprobs = result |
| 45 | + |
| 46 | + # Ensure the return type of logprobs is accurate |
| 47 | + assert isinstance( |
| 48 | + prompt_logprobs, FlattenLogprobs if flatten_logprobs else list |
| 49 | + ) |
| 50 | + assert isinstance( |
| 51 | + decode_logprobs, FlattenLogprobs if flatten_logprobs else list |
| 52 | + ) |
| 53 | + |
| 54 | + ######################## |
| 55 | + # Check prompt logprobs |
| 56 | + ######################## |
| 57 | + assert len(prompt_tokens) == len(prompt_logprobs) |
| 58 | + # No logprob for first prompt token |
| 59 | + assert not prompt_logprobs[0] |
| 60 | + for position, (token, logprobs) in enumerate( |
| 61 | + zip(prompt_tokens[1:], prompt_logprobs[1:]), start=1 |
| 62 | + ): |
| 63 | + # Ensure logprobs of prompt token is always returned |
| 64 | + logprob = logprobs.get(token) |
| 65 | + assert logprob is not None |
| 66 | + assert logprob.rank >= 1 |
| 67 | + # Ensure # of returned logprobs should be |
| 68 | + # either NUM_PROMPT_LOGPROBS or NUM_PROMPT_LOGPROBS+1 |
| 69 | + assert NUM_PROMPT_LOGPROBS <= len(logprobs) <= NUM_PROMPT_LOGPROBS + 1 |
| 70 | + # Ensure top NUM_PROMPT_LOGPROBS is always extracted |
| 71 | + assert set(range(1, NUM_PROMPT_LOGPROBS + 1)).issubset( |
| 72 | + {logprob.rank for logprob in logprobs.values()} |
| 73 | + ) |
| 74 | + |
| 75 | + ######################## |
| 76 | + # Check sample logprobs |
| 77 | + ######################## |
| 78 | + assert len(decode_tokens) == len(decode_logprobs) |
| 79 | + for position, (token, logprobs) in enumerate( |
| 80 | + zip(decode_tokens, decode_logprobs) |
| 81 | + ): |
| 82 | + # Ensure logprobs of chosen token is always returned |
| 83 | + logprob = logprobs.get(token) |
| 84 | + assert logprob is not None |
| 85 | + if greedy: |
| 86 | + # For greedy sampling, all chosen logprob should be top ranked |
| 87 | + assert logprob.rank == 1 |
| 88 | + else: |
| 89 | + assert logprob.rank >= 1 |
| 90 | + # Ensure # of returned logprobs should be |
| 91 | + # either NUM_TOP_LOGPROBS or NUM_TOP_LOGPROBS+1 |
| 92 | + assert NUM_TOP_LOGPROBS <= len(logprobs) <= NUM_TOP_LOGPROBS + 1 |
| 93 | + # Ensure top NUM_TOP_LOGPROBS logprobs is always extracted |
| 94 | + assert set(range(1, NUM_TOP_LOGPROBS + 1)).issubset( |
| 95 | + {logprob.rank for logprob in logprobs.values()} |
| 96 | + ) |
0 commit comments