Add pydocstyle to the ruff rules (#52136)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Signed-off-by: Harry Mellor <hej.mellor@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
804e5377aa
commit
8c1557a79c
@@ -1,7 +1,8 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Test that batched_count_greater_than does not trigger 0/1 specialization
|
||||
recompiles when batch_size varies."""
|
||||
recompiles when batch_size varies.
|
||||
"""
|
||||
|
||||
import torch
|
||||
|
||||
|
||||
@@ -127,6 +127,7 @@ def _repeat_logprob_config(
|
||||
`logprob_prompt_logprob_list` enough times to match the
|
||||
number of `test_prompts`, or else is truncated to match
|
||||
the number of `test_prompts`
|
||||
|
||||
"""
|
||||
num_test_prompts = len(test_prompts)
|
||||
# Make sure there is a logprobs configuration for each test prompt
|
||||
@@ -312,7 +313,7 @@ def test_get_logprobs_and_prompt_logprobs(
|
||||
temperature: float,
|
||||
example_prompts: list[str],
|
||||
) -> None:
|
||||
"""Test V1 Engine logprobs & prompt logprobs
|
||||
"""Test V1 Engine logprobs & prompt logprobs.
|
||||
|
||||
Exercise a variety of combinations of `logprobs` and `prompt_logprobs`
|
||||
settings and validate that
|
||||
@@ -337,6 +338,7 @@ def test_get_logprobs_and_prompt_logprobs(
|
||||
batch_logprobs_composition: logprobs configuration for test batch
|
||||
temperature: "temperature" sampling parameter
|
||||
example_prompts: example prompt fixture
|
||||
|
||||
"""
|
||||
vllm_config = vllm_model.llm.llm_engine.vllm_config
|
||||
do_apc = vllm_config.cache_config.enable_prefix_caching
|
||||
@@ -389,7 +391,7 @@ def test_get_logprobs_and_prompt_logprobs(
|
||||
|
||||
|
||||
def test_max_logprobs():
|
||||
"""vLLM v1 engine should fail a request with `logprobs > max_logprobs`
|
||||
"""VLLM v1 engine should fail a request with `logprobs > max_logprobs`
|
||||
Should also fail for `prompt_logprobs > max_logprobs`
|
||||
APC should not matter as this test checks basic request validation.
|
||||
"""
|
||||
@@ -431,11 +433,12 @@ def test_logprob_token_ids_validate_vocab_bounds_invalid(token_ids: list[int]):
|
||||
|
||||
|
||||
def test_none_logprobs(vllm_model, example_prompts):
|
||||
"""Engine should return `logprobs` and `prompt_logprobs` as `None`
|
||||
"""Engine should return `logprobs` and `prompt_logprobs` as `None`.
|
||||
|
||||
Args:
|
||||
vllm_model: vLLM model fixture
|
||||
example_prompts: list of example prompts (test fixture)
|
||||
|
||||
"""
|
||||
max_tokens = 5
|
||||
|
||||
@@ -459,11 +462,12 @@ def test_none_logprobs(vllm_model, example_prompts):
|
||||
|
||||
|
||||
def test_zero_logprobs(vllm_model, example_prompts):
|
||||
"""Engine should return sampled token and prompt token logprobs
|
||||
"""Engine should return sampled token and prompt token logprobs.
|
||||
|
||||
Args:
|
||||
vllm_model: vLLM model fixture
|
||||
example_prompts: list of example prompts (test fixture)
|
||||
|
||||
"""
|
||||
max_tokens = 5
|
||||
|
||||
@@ -491,10 +495,11 @@ def test_zero_logprobs(vllm_model, example_prompts):
|
||||
|
||||
|
||||
def test_all_logprobs(example_prompts):
|
||||
"""Engine should return all vocabulary logprobs and prompt logprobs
|
||||
"""Engine should return all vocabulary logprobs and prompt logprobs.
|
||||
|
||||
Args:
|
||||
example_prompts: list of example prompts (test fixture)
|
||||
|
||||
"""
|
||||
with VllmRunner(
|
||||
"facebook/opt-125m",
|
||||
@@ -568,7 +573,8 @@ def test_logprobs_mode(logprobs_mode: LogprobsMode):
|
||||
def test_prompt_logprobs_mode():
|
||||
"""prompt_logprobs must respect logprobs_mode: *_logits and *_logprobs
|
||||
must return different values. Prompt tokens skip sampling processors,
|
||||
so processed_* == raw_* on the prompt side."""
|
||||
so processed_* == raw_* on the prompt side.
|
||||
"""
|
||||
from vllm import LLM
|
||||
|
||||
values: dict[str, float] = {}
|
||||
@@ -1127,6 +1133,7 @@ def test_spec_decode_logprobs(
|
||||
logprobs_mode: logprobs mode.
|
||||
model_setup: Tuple of (method, base model name,
|
||||
speculative_config dict, top_logprobs).
|
||||
|
||||
"""
|
||||
from vllm import LLM
|
||||
|
||||
@@ -1224,7 +1231,6 @@ def test_prompt_logprobs_with_chunking_and_preemption():
|
||||
This test ensures that the num_prompt_logprobs tracking persists
|
||||
across preemptions and prefill chunks.
|
||||
"""
|
||||
|
||||
# Create prompts that will trigger chunking and preemption
|
||||
prompts = [
|
||||
"The following numbers of the sequence "
|
||||
|
||||
@@ -54,7 +54,8 @@ def create_logits_tensor(
|
||||
token_idx_to_override: int | None = None,
|
||||
) -> torch.Tensor:
|
||||
"""Helper function to create logits tensor that
|
||||
will produce desired token ids on argmax"""
|
||||
will produce desired token ids on argmax
|
||||
"""
|
||||
token_ids = [tokens[:-1] for tokens in output_token_ids]
|
||||
num_total_tokens = sum(len(tokens) for tokens in token_ids)
|
||||
logits = torch.full((num_total_tokens, vocab_size), -100.0, device=DEVICE_TYPE)
|
||||
@@ -131,7 +132,7 @@ def create_sampling_metadata(
|
||||
|
||||
########################### Tests for Greedy Sampling ###################
|
||||
def test_perfect_match(rejection_sampler):
|
||||
"""Test when output tokens perfectly match speculated tokens"""
|
||||
"""Test when output tokens perfectly match speculated tokens."""
|
||||
spec_tokens = [[1, 2, 3]]
|
||||
output_tokens = [[1, 2, 3, 4]] # 4 is the bonus token
|
||||
|
||||
@@ -152,7 +153,7 @@ def test_perfect_match(rejection_sampler):
|
||||
|
||||
|
||||
def test_early_mismatch(rejection_sampler):
|
||||
"""Test when there's an early mismatch in tokens"""
|
||||
"""Test when there's an early mismatch in tokens."""
|
||||
spec_tokens = [[1, 2, 3]]
|
||||
output_tokens = [[1, 5, 3, 4]] # Mismatch at position 1
|
||||
|
||||
@@ -177,7 +178,7 @@ def test_early_mismatch(rejection_sampler):
|
||||
|
||||
|
||||
def test_multiple_sequences(rejection_sampler):
|
||||
"""Test handling multiple sequences of speculated tokens"""
|
||||
"""Test handling multiple sequences of speculated tokens."""
|
||||
spec_tokens = [[1, 2], [3]]
|
||||
output_tokens = [[1, 2, 5], [3, 4]] # Two sequences with bonus tokens 5 and 4
|
||||
|
||||
@@ -202,7 +203,7 @@ def test_multiple_sequences(rejection_sampler):
|
||||
|
||||
|
||||
def test_single_token_sequence(rejection_sampler):
|
||||
"""Test handling sequences with single token"""
|
||||
"""Test handling sequences with single token."""
|
||||
spec_tokens = [[1]]
|
||||
output_tokens = [[1, 2]] # Single token with bonus token 2
|
||||
|
||||
@@ -223,7 +224,7 @@ def test_single_token_sequence(rejection_sampler):
|
||||
|
||||
|
||||
def test_empty_sequence(rejection_sampler):
|
||||
"""Test handling empty sequence of speculated tokens"""
|
||||
"""Test handling empty sequence of speculated tokens."""
|
||||
spec_tokens: list[list[int]] = [[]]
|
||||
output_tokens = [[5]] # Just the bonus token
|
||||
|
||||
@@ -244,7 +245,7 @@ def test_empty_sequence(rejection_sampler):
|
||||
|
||||
|
||||
def test_multiple_mismatches(rejection_sampler):
|
||||
"""Test handling multiple sequences with mismatches"""
|
||||
"""Test handling multiple sequences with mismatches."""
|
||||
spec_tokens = [[1, 2, 3], [4, 5, 6]]
|
||||
output_tokens = [[1, 2, 7, 6], [4, 8, 6, 9]] # Mismatches in both sequences
|
||||
|
||||
@@ -286,7 +287,7 @@ def test_multiple_mismatches(rejection_sampler):
|
||||
],
|
||||
)
|
||||
def test_parametrized_cases(rejection_sampler, spec_tokens, output_tokens, expected):
|
||||
"""Parametrized test for various matching scenarios"""
|
||||
"""Parametrized test for various matching scenarios."""
|
||||
metadata = create_sampling_metadata(all_greedy=True)
|
||||
logits = create_logits_tensor(output_tokens)
|
||||
bonus_token_tensor = torch.tensor(
|
||||
@@ -486,6 +487,7 @@ def estimate_rejection_sampling_pdf(
|
||||
|
||||
Returns:
|
||||
Estimated probability distribution of the output tokens.
|
||||
|
||||
"""
|
||||
mock_sampler = Mock(spec=Sampler)
|
||||
mock_sampler.logprobs_mode = "raw_logprobs"
|
||||
@@ -656,7 +658,7 @@ def _test_masked_logits(
|
||||
|
||||
@pytest.mark.parametrize("top_k", [1, 5, 99])
|
||||
def test_top_k(rejection_sampler, top_k):
|
||||
"""Test rejection sampling with top-k sampling"""
|
||||
"""Test rejection sampling with top-k sampling."""
|
||||
vocab_size = 100
|
||||
batch_size = 100
|
||||
num_draft_tokens = 3
|
||||
@@ -699,7 +701,7 @@ def test_top_k(rejection_sampler, top_k):
|
||||
|
||||
@pytest.mark.parametrize("top_p", [0.5, 0.9, 0.99])
|
||||
def test_top_p(rejection_sampler, top_p):
|
||||
"""Test rejection sampling with top-p sampling"""
|
||||
"""Test rejection sampling with top-p sampling."""
|
||||
vocab_size = 100
|
||||
batch_size = 100
|
||||
num_draft_tokens = 3
|
||||
@@ -746,7 +748,7 @@ def test_top_p(rejection_sampler, top_p):
|
||||
|
||||
########################### Tests for Logit Processors ###################
|
||||
def test_frequency_penalties(rejection_sampler):
|
||||
"""Test rejection sampling with frequency penalties"""
|
||||
"""Test rejection sampling with frequency penalties."""
|
||||
spec_tokens = [[1, 1, 1], [], [1, 1, 1]]
|
||||
output_tokens = [[1, 1, 1, 1], [7], [1, 1, 1, 1]] # 1, 7 and 1 are the bonus tokens
|
||||
|
||||
@@ -829,7 +831,7 @@ def test_bad_words(rejection_sampler):
|
||||
|
||||
|
||||
def test_allowed_token_ids(rejection_sampler):
|
||||
"""Test rejection sampling with allowed token ids"""
|
||||
"""Test rejection sampling with allowed token ids."""
|
||||
spec_tokens = [[1, 2, 10], [10, 5, 3], [7, 10, 12]]
|
||||
output_tokens = [[1, 2, 10, 5], [10, 5, 10, 5], [7, 10, 12, 5]]
|
||||
# Not allowed tokens:
|
||||
|
||||
@@ -169,8 +169,7 @@ def _create_default_sampling_metadata(
|
||||
def _create_weighted_output_token_list(
|
||||
batch_size: int, vocab_size: int
|
||||
) -> tuple[list[list[int]], list[list[int]]]:
|
||||
"""
|
||||
Creates an output token list where each token occurs a distinct
|
||||
"""Creates an output token list where each token occurs a distinct
|
||||
number of times.
|
||||
|
||||
For each batch, a random subset of token IDs is selected from the
|
||||
@@ -185,6 +184,7 @@ def _create_weighted_output_token_list(
|
||||
- The second element is a list of distinct token IDs for each
|
||||
batch, ordered by their frequency in the corresponding output
|
||||
list.
|
||||
|
||||
"""
|
||||
output_token_ids: list[list[int]] = []
|
||||
sorted_token_ids_in_output: list[list[int]] = []
|
||||
@@ -206,8 +206,7 @@ def _create_weighted_output_token_list(
|
||||
def test_sampler_presence_penalty(
|
||||
device: str, batch_size: int, presence_penalty: float
|
||||
):
|
||||
"""
|
||||
Test to verify that if presence penalty is enabled then tokens
|
||||
"""Test to verify that if presence penalty is enabled then tokens
|
||||
are penalized as per their presence in the existing output.
|
||||
"""
|
||||
torch.set_default_device(device)
|
||||
@@ -256,8 +255,7 @@ def test_sampler_presence_penalty(
|
||||
def test_sampler_frequency_penalty(
|
||||
device: str, batch_size: int, frequency_penalty: float
|
||||
):
|
||||
"""
|
||||
Test to verify that if frequency penalty is enabled then tokens are
|
||||
"""Test to verify that if frequency penalty is enabled then tokens are
|
||||
penalized as per their frequency of occurrence.
|
||||
"""
|
||||
torch.set_default_device(device)
|
||||
@@ -312,8 +310,7 @@ def test_sampler_frequency_penalty(
|
||||
def test_sampler_repetition_penalty(
|
||||
device: str, batch_size: int, repetition_penalty: float
|
||||
):
|
||||
"""
|
||||
Test to verify that when the repetition penalty is enabled, tokens
|
||||
"""Test to verify that when the repetition penalty is enabled, tokens
|
||||
are penalized based on their presence in the prompt or the existing
|
||||
output.
|
||||
"""
|
||||
@@ -370,8 +367,7 @@ def test_sampler_repetition_penalty(
|
||||
def test_sampler_allowed_token_ids(
|
||||
device: str, batch_size: int, num_allowed_token_ids: int
|
||||
):
|
||||
"""
|
||||
Test to verify that when the repetition penalty is enabled, tokens
|
||||
"""Test to verify that when the repetition penalty is enabled, tokens
|
||||
are penalized based on their presence in the prompt or the existing
|
||||
output.
|
||||
"""
|
||||
@@ -416,8 +412,7 @@ def test_sampler_allowed_token_ids(
|
||||
def test_sampler_bad_words(
|
||||
device: str, batch_size: int, bad_words_lengths: tuple[int, ...]
|
||||
):
|
||||
"""
|
||||
Test to verify that when the bad words restriction is present, tokens
|
||||
"""Test to verify that when the bad words restriction is present, tokens
|
||||
are penalized based on their match with the bad words.
|
||||
"""
|
||||
torch.set_default_device(device)
|
||||
|
||||
@@ -17,7 +17,6 @@ def llm() -> LLM:
|
||||
|
||||
def test_n_gt_1(llm):
|
||||
"""ParallelSampling is supported."""
|
||||
|
||||
params = SamplingParams(n=3)
|
||||
outputs = llm.generate(PROMPT, params)
|
||||
assert len(outputs[0].outputs) == 3
|
||||
@@ -25,7 +24,6 @@ def test_n_gt_1(llm):
|
||||
|
||||
def test_penalties(llm):
|
||||
"""Check that we do not get errors if applied."""
|
||||
|
||||
params = SamplingParams(
|
||||
temperature=1.2,
|
||||
presence_penalty=1.2,
|
||||
@@ -40,7 +38,6 @@ def test_penalties(llm):
|
||||
|
||||
def test_stop(llm):
|
||||
"""Check that we respect the stop words."""
|
||||
|
||||
output = llm.generate(PROMPT, SamplingParams(temperature=0))
|
||||
split_text = output[0].outputs[0].text.split()
|
||||
|
||||
@@ -64,7 +61,6 @@ def test_stop(llm):
|
||||
|
||||
def test_stop_token_ids(llm):
|
||||
"""Check that we respect the stop token ids."""
|
||||
|
||||
output = llm.generate(PROMPT, SamplingParams(temperature=0))
|
||||
|
||||
stop_token_id_0 = output[0].outputs[0].token_ids[5]
|
||||
@@ -83,7 +79,6 @@ def test_stop_token_ids(llm):
|
||||
|
||||
def test_detokenize_false(llm):
|
||||
"""Check that detokenize=False option works."""
|
||||
|
||||
output = llm.generate(PROMPT, SamplingParams(detokenize=False))
|
||||
assert len(output[0].outputs[0].token_ids) > 0
|
||||
assert len(output[0].outputs[0].text) == 0
|
||||
@@ -106,7 +101,6 @@ def test_detokenize_false(llm):
|
||||
|
||||
def test_bad_words(llm):
|
||||
"""Check that we respect bad words."""
|
||||
|
||||
tokenizer = llm.get_tokenizer()
|
||||
|
||||
def contains_bad_word(text: str, tokens: list[int], bad_word: str) -> bool:
|
||||
@@ -147,7 +141,6 @@ def test_bad_words(llm):
|
||||
|
||||
def test_allowed_token_ids(llm):
|
||||
"""Check that we can use allowed_token_ids."""
|
||||
|
||||
TOKEN_ID = 10
|
||||
allowed_token_ids = [TOKEN_ID]
|
||||
output = llm.generate(PROMPT, SamplingParams(allowed_token_ids=allowed_token_ids))
|
||||
@@ -176,7 +169,6 @@ def test_allowed_token_ids(llm):
|
||||
|
||||
def test_seed(llm):
|
||||
"""Check that seed impacts randomness."""
|
||||
|
||||
out_1 = llm.generate(PROMPT, SamplingParams(seed=42))
|
||||
out_2 = llm.generate(PROMPT, SamplingParams(seed=42))
|
||||
out_3 = llm.generate(PROMPT, SamplingParams(seed=43))
|
||||
|
||||
@@ -50,8 +50,7 @@ def _seed_default_generator(seed: int) -> None:
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def reset_default_device():
|
||||
"""
|
||||
Explicitly set the default device, which can affect subsequent tests.
|
||||
"""Explicitly set the default device, which can affect subsequent tests.
|
||||
Adding this fixture helps avoid this problem.
|
||||
"""
|
||||
original_device = torch.get_default_device()
|
||||
@@ -171,8 +170,7 @@ def test_topk_impl_equivalence():
|
||||
"interface differences between Python and FlashInfer implementations"
|
||||
)
|
||||
def test_flashinfer_sampler():
|
||||
"""
|
||||
This test verifies that the FlashInfer top-k and top-p sampling
|
||||
"""This test verifies that the FlashInfer top-k and top-p sampling
|
||||
implementation produces the same results as the Python implementation.
|
||||
|
||||
NOTE: FlashInfer did not directly expose an interface for fused top-k and
|
||||
@@ -886,7 +884,8 @@ class TestFlashInferTopkToppRobustness:
|
||||
|
||||
def _make_logits(self, pattern: str) -> torch.Tensor:
|
||||
"""Build (BATCH, VOCAB) logits with `pattern` applied to row 0
|
||||
(rows 1..B-1 stay clean so we can detect cross-row corruption)."""
|
||||
(rows 1..B-1 stay clean so we can detect cross-row corruption).
|
||||
"""
|
||||
logits = (
|
||||
torch.randn(
|
||||
self.BATCH,
|
||||
|
||||
@@ -15,7 +15,7 @@ from vllm.v1.sample.metadata import SamplingMetadata
|
||||
|
||||
|
||||
class BatchLogprobsComposition(Enum):
|
||||
"""Types of logprobs configs to include in test batch"""
|
||||
"""Types of logprobs configs to include in test batch."""
|
||||
|
||||
NONE = 0
|
||||
SAMPLE = 1
|
||||
@@ -29,7 +29,7 @@ BatchLogprobsSpecType = list[tuple[int | None, int | None]]
|
||||
def get_test_batch(
|
||||
batch_logprobs_composition: BatchLogprobsComposition,
|
||||
) -> BatchLogprobsSpecType:
|
||||
"""Generate logprobs configs for a batch of requests
|
||||
"""Generate logprobs configs for a batch of requests.
|
||||
|
||||
A given request's logprobs configuration is (1) num_sample_logprobs and (2)
|
||||
num_prompt_logprobs. The batch logprobs configuration is the list of request
|
||||
@@ -53,9 +53,9 @@ def get_test_batch(
|
||||
batch_logprobs_composition: types of logprobs configs to include in batch
|
||||
|
||||
Returns:
|
||||
|
||||
list of (Optional[num_sample_logprobs], Optional[num_prompt_logprobs])
|
||||
tuples
|
||||
|
||||
"""
|
||||
if batch_logprobs_composition == BatchLogprobsComposition.NONE:
|
||||
# No requests with sample or prompt logprobs
|
||||
@@ -101,7 +101,7 @@ def assert_incr_detok_str_matches_non_incr_detok_str(
|
||||
non_incremental_detokenization_str: str,
|
||||
msg: str,
|
||||
) -> None:
|
||||
"""Compare incrementally detok. text to non-incrementally detok. text
|
||||
"""Compare incrementally detok. text to non-incrementally detok. text.
|
||||
|
||||
Fail if the strings mismatch after non-alphanumeric characters are stripped
|
||||
out.
|
||||
@@ -120,6 +120,7 @@ def assert_incr_detok_str_matches_non_incr_detok_str(
|
||||
non_incremental_detokenization_str: non-incrementally-detokenized logprob
|
||||
tokens
|
||||
msg: error message if `assert` fails
|
||||
|
||||
"""
|
||||
rgx = r"[^a-zA-Z0-9]+"
|
||||
assert re.sub(rgx, "", incremental_detokenization_str) == re.sub(
|
||||
@@ -128,13 +129,14 @@ def assert_incr_detok_str_matches_non_incr_detok_str(
|
||||
|
||||
|
||||
def compute_correct_cumulative_logprob(completion_output: CompletionOutput) -> float:
|
||||
"""Compute known-good value for evaluating cumulative logprob
|
||||
"""Compute known-good value for evaluating cumulative logprob.
|
||||
|
||||
Args:
|
||||
completion_output: completion output from engine
|
||||
|
||||
Returns:
|
||||
Known-good cumulative logprob value
|
||||
|
||||
"""
|
||||
token_ids = completion_output.token_ids
|
||||
logprobs = completion_output.logprobs
|
||||
@@ -170,7 +172,7 @@ def create_prompt_tokens_tensor(
|
||||
|
||||
|
||||
class LogitsprocsTestFakes(NamedTuple):
|
||||
"""Wraps fake data structures to support testing"""
|
||||
"""Wraps fake data structures to support testing."""
|
||||
|
||||
logits: torch.Tensor
|
||||
sampling_metadata: SamplingMetadata
|
||||
@@ -186,6 +188,7 @@ class LogitsprocsTestFakes(NamedTuple):
|
||||
|
||||
Returns:
|
||||
Iterator over logits processors
|
||||
|
||||
"""
|
||||
return (
|
||||
lp for lp in self.sampling_metadata.logitsprocs.all if isinstance(lp, cls)
|
||||
@@ -201,7 +204,8 @@ def fake_update_logitsprocs_state(
|
||||
batch_update: BatchUpdate | None,
|
||||
) -> None:
|
||||
"""Imitate logits processors persistent batch state update
|
||||
in engine core"""
|
||||
in engine core
|
||||
"""
|
||||
for logitproc in test_fakes.get_logitsprocs():
|
||||
logitproc.update_state(batch_update)
|
||||
holder = test_fakes.sampling_metadata.thinking_budget_state_holder
|
||||
|
||||
Reference in New Issue
Block a user