Convert formatting to use ruff instead of yapf + isort (#26247)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -36,19 +36,21 @@ def test_beam_search_single_input(
|
||||
) -> None:
|
||||
example_prompts = example_prompts[:1]
|
||||
with hf_runner(model, dtype=dtype) as hf_model:
|
||||
hf_outputs = hf_model.generate_beam_search(example_prompts, beam_width,
|
||||
max_tokens)
|
||||
hf_outputs = hf_model.generate_beam_search(
|
||||
example_prompts, beam_width, max_tokens
|
||||
)
|
||||
|
||||
with vllm_runner(model, dtype=dtype) as vllm_model:
|
||||
vllm_outputs = vllm_model.generate_beam_search(example_prompts,
|
||||
beam_width, max_tokens)
|
||||
vllm_outputs = vllm_model.generate_beam_search(
|
||||
example_prompts, beam_width, max_tokens
|
||||
)
|
||||
|
||||
for i in range(len(example_prompts)):
|
||||
hf_output_ids, hf_output_texts = hf_outputs[i]
|
||||
vllm_output_ids, vllm_output_texts = vllm_outputs[i]
|
||||
for j, (hf_text,
|
||||
vllm_text) in enumerate(zip(hf_output_texts,
|
||||
vllm_output_texts)):
|
||||
for j, (hf_text, vllm_text) in enumerate(
|
||||
zip(hf_output_texts, vllm_output_texts)
|
||||
):
|
||||
print(f">>>{j}-th hf output:")
|
||||
print(hf_text)
|
||||
print(f">>>{j}-th vllm output:")
|
||||
@@ -56,8 +58,8 @@ def test_beam_search_single_input(
|
||||
assert len(hf_output_ids) == len(vllm_output_ids)
|
||||
for j in range(len(hf_output_ids)):
|
||||
assert hf_output_ids[j] == vllm_output_ids[j], (
|
||||
f"Test{i} output{j}:\nHF: {hf_output_ids}\n"
|
||||
f"vLLM: {vllm_output_ids}")
|
||||
f"Test{i} output{j}:\nHF: {hf_output_ids}\nvLLM: {vllm_output_ids}"
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.skip_v1 # FIXME: This fails on V1 right now.
|
||||
@@ -76,30 +78,29 @@ def test_beam_search_with_concurrency_limit(
|
||||
) -> None:
|
||||
# example_prompts[1]&[3]&[7] fails due to unknown reason even without
|
||||
# concurrency limit. skip them for now.
|
||||
example_prompts = (example_prompts[:8])
|
||||
example_prompts = example_prompts[:8]
|
||||
concurrency_limit = 2
|
||||
assert len(example_prompts) > concurrency_limit
|
||||
with vllm_runner(model, dtype=dtype) as vllm_model:
|
||||
outputs_with_limit = vllm_model.generate_beam_search(
|
||||
example_prompts,
|
||||
beam_width,
|
||||
max_tokens,
|
||||
concurrency_limit=concurrency_limit)
|
||||
example_prompts, beam_width, max_tokens, concurrency_limit=concurrency_limit
|
||||
)
|
||||
outputs_without_limit = []
|
||||
|
||||
for i in range(0, len(example_prompts), concurrency_limit):
|
||||
outputs_without_limit.extend(
|
||||
vllm_model.generate_beam_search(
|
||||
example_prompts[i:i + concurrency_limit], beam_width,
|
||||
max_tokens))
|
||||
example_prompts[i : i + concurrency_limit], beam_width, max_tokens
|
||||
)
|
||||
)
|
||||
|
||||
correct = True
|
||||
for i in range(len(example_prompts)):
|
||||
output_ids_with_limit, output_texts_with_limit = outputs_with_limit[i]
|
||||
output_ids_without_limit, output_texts_without_limit = (
|
||||
outputs_without_limit[i])
|
||||
output_ids_without_limit, output_texts_without_limit = outputs_without_limit[i]
|
||||
for j, (text_with_limit, text_without_limit) in enumerate(
|
||||
zip(output_texts_with_limit, output_texts_without_limit)):
|
||||
zip(output_texts_with_limit, output_texts_without_limit)
|
||||
):
|
||||
print(f">>>{j}-th with limit output:")
|
||||
print(text_with_limit)
|
||||
print(f">>>{j}-th without limit output:")
|
||||
@@ -107,8 +108,10 @@ def test_beam_search_with_concurrency_limit(
|
||||
assert len(output_ids_with_limit) == len(output_ids_without_limit)
|
||||
for j in range(len(output_ids_with_limit)):
|
||||
if output_ids_with_limit[j] != output_ids_without_limit[j]:
|
||||
print(f"Test{i} output{j}:\n+limit: {output_ids_with_limit}\n"
|
||||
f"-limit: {output_ids_without_limit}")
|
||||
print(
|
||||
f"Test{i} output{j}:\n+limit: {output_ids_with_limit}\n"
|
||||
f"-limit: {output_ids_without_limit}"
|
||||
)
|
||||
correct = False
|
||||
assert correct
|
||||
|
||||
@@ -131,11 +134,10 @@ def test_beam_search_passes_multimodal_data(
|
||||
model = "Qwen/Qwen2-Audio-7B-Instruct"
|
||||
audio_seq = "<|audio_bos|><|AUDIO|><|audio_eos|>"
|
||||
prompts = [
|
||||
f"<|im_start|>user\n{audio_seq}Can you transcribe this?<|im_end|>\n<|im_start|>assistant\n" #noqa: E501
|
||||
f"<|im_start|>user\n{audio_seq}Can you transcribe this?<|im_end|>\n<|im_start|>assistant\n" # noqa: E501
|
||||
]
|
||||
|
||||
with hf_runner(model, dtype=dtype,
|
||||
auto_cls=AutoModelForSeq2SeqLM) as hf_model:
|
||||
with hf_runner(model, dtype=dtype, auto_cls=AutoModelForSeq2SeqLM) as hf_model:
|
||||
audio_token_id = hf_model.config.audio_token_index
|
||||
eos_token_id = hf_model.tokenizer.eos_token_id # <|im_end|>
|
||||
hf_outputs = hf_model.generate_beam_search(
|
||||
@@ -153,17 +155,15 @@ def test_beam_search_passes_multimodal_data(
|
||||
audios=audios,
|
||||
)
|
||||
|
||||
seq_with_no_audio_toks = lambda seq: [
|
||||
tok for tok in seq if tok != audio_token_id
|
||||
]
|
||||
seq_with_no_audio_toks = lambda seq: [tok for tok in seq if tok != audio_token_id]
|
||||
|
||||
for i in range(len(prompts)):
|
||||
hf_output_ids, hf_output_texts = hf_outputs[i]
|
||||
vllm_output_ids, vllm_output_texts = vllm_outputs[i]
|
||||
|
||||
for j, (hf_text,
|
||||
vllm_text) in enumerate(zip(hf_output_texts,
|
||||
vllm_output_texts)):
|
||||
for j, (hf_text, vllm_text) in enumerate(
|
||||
zip(hf_output_texts, vllm_output_texts)
|
||||
):
|
||||
print(f">>>{j}-th hf output [NOTE: special tokens are filtered]:")
|
||||
print(hf_text)
|
||||
print(f">>>{j}-th vllm output:")
|
||||
@@ -176,12 +176,10 @@ def test_beam_search_passes_multimodal_data(
|
||||
# token to match features, while the vLLM helper maintains the
|
||||
# single audio token in the input text
|
||||
filtered_hf_output_ids = seq_with_no_audio_toks(hf_output_ids[j])
|
||||
filtered_vllm_output_ids = seq_with_no_audio_toks(
|
||||
vllm_output_ids[j])
|
||||
filtered_vllm_output_ids = seq_with_no_audio_toks(vllm_output_ids[j])
|
||||
|
||||
# HF output IDs may contain the end of sequence
|
||||
if len(filtered_hf_output_ids
|
||||
) == len(filtered_vllm_output_ids) + 1:
|
||||
if len(filtered_hf_output_ids) == len(filtered_vllm_output_ids) + 1:
|
||||
assert filtered_hf_output_ids[-1] == eos_token_id
|
||||
filtered_hf_output_ids = filtered_hf_output_ids[:-1]
|
||||
|
||||
|
||||
@@ -25,11 +25,11 @@ def test_ignore_eos(
|
||||
max_tokens: int,
|
||||
) -> None:
|
||||
with vllm_runner(model, dtype=dtype) as vllm_model:
|
||||
sampling_params = SamplingParams(max_tokens=max_tokens,
|
||||
ignore_eos=True)
|
||||
sampling_params = SamplingParams(max_tokens=max_tokens, ignore_eos=True)
|
||||
|
||||
for prompt in example_prompts:
|
||||
ignore_eos_output = vllm_model.llm.generate(
|
||||
prompt, sampling_params=sampling_params)
|
||||
prompt, sampling_params=sampling_params
|
||||
)
|
||||
output_length = len(ignore_eos_output[0].outputs[0].token_ids)
|
||||
assert output_length == max_tokens
|
||||
|
||||
@@ -5,6 +5,7 @@
|
||||
Run `pytest tests/samplers/test_no_bad_words.py`.
|
||||
|
||||
"""
|
||||
|
||||
from typing import Optional
|
||||
|
||||
import pytest
|
||||
@@ -16,7 +17,7 @@ from vllm import LLM, SamplingParams
|
||||
@pytest.fixture(autouse=True)
|
||||
def v1(monkeypatch):
|
||||
"""Only run on vLLM v1."""
|
||||
monkeypatch.setenv('VLLM_USE_V1', '1')
|
||||
monkeypatch.setenv("VLLM_USE_V1", "1")
|
||||
|
||||
|
||||
def _generate(
|
||||
@@ -49,25 +50,24 @@ class TestOneTokenBadWord:
|
||||
TARGET_TOKEN = "you"
|
||||
|
||||
def setup_method(self, method):
|
||||
self.tokenizer = AutoTokenizer.from_pretrained(self.MODEL,
|
||||
add_prefix_space=True)
|
||||
self.tokenizer = AutoTokenizer.from_pretrained(
|
||||
self.MODEL, add_prefix_space=True
|
||||
)
|
||||
|
||||
self.num_prompt_tokens = len(self._encode(self.PROMPT))
|
||||
self.target_token_id = self._encode(self.TARGET_TOKEN,
|
||||
add_special_tokens=False)[0]
|
||||
self.target_token_id = self._encode(
|
||||
self.TARGET_TOKEN, add_special_tokens=False
|
||||
)[0]
|
||||
|
||||
def test_one_token_bad_word(self, vllm_runner):
|
||||
with vllm_runner(self.MODEL) as llm:
|
||||
output_token_ids = self._generate(llm)
|
||||
assert output_token_ids[0] == self.target_token_id
|
||||
|
||||
output_token_ids = self._generate(llm,
|
||||
bad_words=[self.TARGET_TOKEN])
|
||||
output_token_ids = self._generate(llm, bad_words=[self.TARGET_TOKEN])
|
||||
assert self.target_token_id not in output_token_ids
|
||||
|
||||
def _generate(self,
|
||||
llm: LLM,
|
||||
bad_words: Optional[list[str]] = None) -> list[int]:
|
||||
def _generate(self, llm: LLM, bad_words: Optional[list[str]] = None) -> list[int]:
|
||||
return _generate(
|
||||
llm=llm,
|
||||
prompt=self.PROMPT,
|
||||
@@ -75,11 +75,8 @@ class TestOneTokenBadWord:
|
||||
bad_words=bad_words,
|
||||
)
|
||||
|
||||
def _encode(self,
|
||||
prompt: str,
|
||||
add_special_tokens: bool = True) -> list[int]:
|
||||
return self.tokenizer(prompt,
|
||||
add_special_tokens=add_special_tokens).input_ids
|
||||
def _encode(self, prompt: str, add_special_tokens: bool = True) -> list[int]:
|
||||
return self.tokenizer(prompt, add_special_tokens=add_special_tokens).input_ids
|
||||
|
||||
|
||||
class TestTwoTokenBadWord:
|
||||
@@ -92,72 +89,80 @@ class TestTwoTokenBadWord:
|
||||
NEIGHBOUR_TOKEN2 = "older"
|
||||
|
||||
def setup_method(self, method):
|
||||
self.tokenizer = AutoTokenizer.from_pretrained(self.MODEL,
|
||||
add_prefix_space=True)
|
||||
self.tokenizer = AutoTokenizer.from_pretrained(
|
||||
self.MODEL, add_prefix_space=True
|
||||
)
|
||||
|
||||
self.num_prompt_tokens = len(self._encode(self.PROMPT))
|
||||
self.target_token_id1 = self._encode(self.TARGET_TOKEN1,
|
||||
add_special_tokens=False)[0]
|
||||
self.target_token_id2 = self._encode(self.TARGET_TOKEN2,
|
||||
add_special_tokens=False)[0]
|
||||
self.neighbour_token_id2 = self._encode(self.NEIGHBOUR_TOKEN2,
|
||||
add_special_tokens=False)[0]
|
||||
self.target_token_id1 = self._encode(
|
||||
self.TARGET_TOKEN1, add_special_tokens=False
|
||||
)[0]
|
||||
self.target_token_id2 = self._encode(
|
||||
self.TARGET_TOKEN2, add_special_tokens=False
|
||||
)[0]
|
||||
self.neighbour_token_id2 = self._encode(
|
||||
self.NEIGHBOUR_TOKEN2, add_special_tokens=False
|
||||
)[0]
|
||||
|
||||
def test_two_token_bad_word(self, vllm_runner):
|
||||
with vllm_runner(self.MODEL, dtype="half") as llm:
|
||||
output_token_ids = self._generate(llm)
|
||||
assert output_token_ids[:2] == [
|
||||
self.target_token_id1, self.target_token_id2
|
||||
self.target_token_id1,
|
||||
self.target_token_id2,
|
||||
]
|
||||
|
||||
output_token_ids = self._generate(llm,
|
||||
bad_words=[self.TARGET_TOKEN1])
|
||||
output_token_ids = self._generate(llm, bad_words=[self.TARGET_TOKEN1])
|
||||
assert self.target_token_id1 not in output_token_ids
|
||||
|
||||
output_token_ids = self._generate(llm,
|
||||
bad_words=[self.TARGET_TOKEN2])
|
||||
output_token_ids = self._generate(llm, bad_words=[self.TARGET_TOKEN2])
|
||||
assert output_token_ids[0] == self.target_token_id1
|
||||
assert self.target_token_id2 not in output_token_ids
|
||||
|
||||
output_token_ids = self._generate(
|
||||
llm, bad_words=[f'{self.TARGET_TOKEN1} {self.TARGET_TOKEN2}'])
|
||||
llm, bad_words=[f"{self.TARGET_TOKEN1} {self.TARGET_TOKEN2}"]
|
||||
)
|
||||
assert output_token_ids[0] == self.target_token_id1
|
||||
assert output_token_ids[:2] != [
|
||||
self.target_token_id1, self.target_token_id2
|
||||
self.target_token_id1,
|
||||
self.target_token_id2,
|
||||
]
|
||||
assert not self._contains(
|
||||
output_token_ids,
|
||||
[self.target_token_id1, self.target_token_id2])
|
||||
output_token_ids, [self.target_token_id1, self.target_token_id2]
|
||||
)
|
||||
# Model dependent behaviour
|
||||
assert output_token_ids[:2] == [
|
||||
self.target_token_id1, self.neighbour_token_id2
|
||||
self.target_token_id1,
|
||||
self.neighbour_token_id2,
|
||||
]
|
||||
|
||||
output_token_ids = self._generate(
|
||||
llm,
|
||||
bad_words=[
|
||||
f'{self.TARGET_TOKEN1} {self.TARGET_TOKEN2}',
|
||||
f'{self.TARGET_TOKEN1} {self.NEIGHBOUR_TOKEN2}'
|
||||
])
|
||||
f"{self.TARGET_TOKEN1} {self.TARGET_TOKEN2}",
|
||||
f"{self.TARGET_TOKEN1} {self.NEIGHBOUR_TOKEN2}",
|
||||
],
|
||||
)
|
||||
assert output_token_ids[0] == self.target_token_id1
|
||||
assert output_token_ids[:2] != [
|
||||
self.target_token_id1, self.target_token_id2
|
||||
self.target_token_id1,
|
||||
self.target_token_id2,
|
||||
]
|
||||
assert not self._contains(
|
||||
output_token_ids,
|
||||
[self.target_token_id1, self.target_token_id2])
|
||||
output_token_ids, [self.target_token_id1, self.target_token_id2]
|
||||
)
|
||||
assert output_token_ids[:2] != [
|
||||
self.target_token_id1, self.neighbour_token_id2
|
||||
self.target_token_id1,
|
||||
self.neighbour_token_id2,
|
||||
]
|
||||
assert not self._contains(
|
||||
output_token_ids,
|
||||
[self.target_token_id1, self.neighbour_token_id2])
|
||||
assert ((self.target_token_id2 in output_token_ids)
|
||||
or (self.neighbour_token_id2 in output_token_ids))
|
||||
output_token_ids, [self.target_token_id1, self.neighbour_token_id2]
|
||||
)
|
||||
assert (self.target_token_id2 in output_token_ids) or (
|
||||
self.neighbour_token_id2 in output_token_ids
|
||||
)
|
||||
|
||||
def _generate(self,
|
||||
llm: LLM,
|
||||
bad_words: Optional[list[str]] = None) -> list[int]:
|
||||
def _generate(self, llm: LLM, bad_words: Optional[list[str]] = None) -> list[int]:
|
||||
return _generate(
|
||||
llm=llm,
|
||||
prompt=self.PROMPT,
|
||||
@@ -187,8 +192,5 @@ class TestTwoTokenBadWord:
|
||||
|
||||
return False
|
||||
|
||||
def _encode(self,
|
||||
prompt: str,
|
||||
add_special_tokens: bool = True) -> list[int]:
|
||||
return self.tokenizer(prompt,
|
||||
add_special_tokens=add_special_tokens).input_ids
|
||||
def _encode(self, prompt: str, add_special_tokens: bool = True) -> list[int]:
|
||||
return self.tokenizer(prompt, add_special_tokens=add_special_tokens).input_ids
|
||||
|
||||
@@ -20,25 +20,27 @@ def test_ranks(
|
||||
num_top_logprobs = 5
|
||||
num_prompt_logprobs = 5
|
||||
|
||||
with vllm_runner(model, dtype=dtype,
|
||||
max_logprobs=num_top_logprobs) as vllm_model:
|
||||
|
||||
with vllm_runner(model, dtype=dtype, max_logprobs=num_top_logprobs) as vllm_model:
|
||||
## Test greedy logprobs ranks
|
||||
vllm_sampling_params = SamplingParams(
|
||||
temperature=0.0,
|
||||
top_p=1.0,
|
||||
max_tokens=max_tokens,
|
||||
logprobs=num_top_logprobs,
|
||||
prompt_logprobs=num_prompt_logprobs)
|
||||
vllm_results = vllm_model.generate_w_logprobs(example_prompts,
|
||||
vllm_sampling_params)
|
||||
prompt_logprobs=num_prompt_logprobs,
|
||||
)
|
||||
vllm_results = vllm_model.generate_w_logprobs(
|
||||
example_prompts, vllm_sampling_params
|
||||
)
|
||||
|
||||
## Test non-greedy logprobs ranks
|
||||
sampling_params = SamplingParams(temperature=1.0,
|
||||
top_p=1.0,
|
||||
max_tokens=max_tokens,
|
||||
logprobs=num_top_logprobs,
|
||||
prompt_logprobs=num_prompt_logprobs)
|
||||
sampling_params = SamplingParams(
|
||||
temperature=1.0,
|
||||
top_p=1.0,
|
||||
max_tokens=max_tokens,
|
||||
logprobs=num_top_logprobs,
|
||||
prompt_logprobs=num_prompt_logprobs,
|
||||
)
|
||||
res = vllm_model.generate_w_logprobs(example_prompts, sampling_params)
|
||||
|
||||
for result in vllm_results:
|
||||
|
||||
Reference in New Issue
Block a user