[V0 Deprecation] Remove V0 Sequence class & Sampler (#25332)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: Woosuk Kwon <woosuk@thinkingmachines.ai>
This commit is contained in:
Woosuk Kwon
2025-09-21 08:52:15 -07:00
committed by GitHub
parent 65a5910ce3
commit 26e673fe93
27 changed files with 69 additions and 3696 deletions
+1 -1
View File
@@ -48,10 +48,10 @@ from vllm.distributed import (cleanup_dist_env_and_memory,
initialize_model_parallel)
from vllm.inputs import TextPrompt
from vllm.logger import init_logger
from vllm.logprobs import Logprob
from vllm.multimodal.utils import fetch_image
from vllm.outputs import RequestOutput
from vllm.sampling_params import BeamSearchParams
from vllm.sequence import Logprob
from vllm.transformers_utils.utils import maybe_model_redirect
from vllm.utils import set_default_torch_num_threads
@@ -7,8 +7,8 @@ from typing import Optional
import pytest
from transformers import AutoModelForSpeechSeq2Seq
from vllm.logprobs import SampleLogprobs
from vllm.lora.request import LoRARequest
from vllm.sequence import SampleLogprobs
from ....conftest import (AudioTestAssets, HfRunner, PromptAudioInput,
VllmRunner)
@@ -12,10 +12,10 @@ from huggingface_hub import snapshot_download
from transformers import AutoTokenizer
from vllm.assets.image import ImageAsset
from vllm.logprobs import SampleLogprobs
from vllm.lora.request import LoRARequest
from vllm.multimodal.image import convert_image_mode, rescale_image_size
from vllm.platforms import current_platform
from vllm.sequence import SampleLogprobs
from ....conftest import (IMAGE_ASSETS, HfRunner, PromptAudioInput,
PromptImageInput, VllmRunner)
@@ -13,8 +13,8 @@ from mistral_common.tokens.tokenizers.multimodal import image_from_chunk
from transformers import AutoProcessor
from vllm import SamplingParams, TextPrompt, TokensPrompt
from vllm.logprobs import Logprob, SampleLogprobs
from vllm.multimodal import MultiModalDataBuiltins
from vllm.sequence import Logprob, SampleLogprobs
from ....utils import VLLM_PATH, large_gpu_test
from ...utils import check_logprobs_close
@@ -19,7 +19,7 @@ from transformers import (AutoConfig, AutoTokenizer, BatchFeature,
GenerationConfig, GenerationMixin)
from transformers.video_utils import VideoMetadata
from vllm.sequence import SampleLogprobs
from vllm.logprobs import SampleLogprobs
from vllm.utils import is_list_of
from .....conftest import HfRunner, ImageAsset, ImageTestAssets
@@ -12,7 +12,7 @@ from transformers import AutoModelForCausalLM
from transformers.models.auto.auto_factory import _BaseAutoModelClass
from vllm.config import RunnerOption
from vllm.sequence import SampleLogprobs
from vllm.logprobs import SampleLogprobs
from vllm.transformers_utils.tokenizer import AnyTokenizer
from .....conftest import (AUDIO_ASSETS, IMAGE_ASSETS, HfRunner, ImageAsset,
+1 -1
View File
@@ -12,7 +12,7 @@ from transformers import PretrainedConfig
from vllm.config import ModelConfig, ModelDType, RunnerOption
from vllm.inputs import InputContext
from vllm.sequence import Logprob, PromptLogprobs, SampleLogprobs
from vllm.logprobs import Logprob, PromptLogprobs, SampleLogprobs
from .registry import HF_EXAMPLE_MODELS
+1 -139
View File
@@ -8,10 +8,7 @@ import pytest
from transformers import (AutoTokenizer, PreTrainedTokenizer,
PreTrainedTokenizerFast)
from vllm.inputs import token_inputs
from vllm.sequence import Logprob, SamplingParams, Sequence, SequenceGroup
from vllm.transformers_utils.detokenizer import Detokenizer
from vllm.transformers_utils.tokenizer import get_tokenizer
from vllm.sampling_params import SamplingParams
from vllm.transformers_utils.tokenizers.mistral import MistralTokenizer
from vllm.v1.engine import EngineCoreRequest
from vllm.v1.engine.detokenizer import (FastIncrementalDetokenizer,
@@ -217,138 +214,3 @@ def test_oov_decode(tokenizer, fast):
assert decoded_text == ''
assert out_ids == [len(tokenizer)]
@pytest.fixture
def detokenizer(tokenizer_name: str) -> Detokenizer:
tokenizer = get_tokenizer(
tokenizer_name,
tokenizer_mode="mistral" if "mistral" in tokenizer_name else "auto",
trust_remote_code=False,
revision=None,
)
return Detokenizer(tokenizer)
@pytest.fixture(name="complete_sequence_token_ids")
def create_complete_sequence_token_ids(complete_sequence: str,
tokenizer) -> list[int]:
return tokenizer(complete_sequence, add_special_tokens=False).input_ids
def create_sequence(prompt_token_ids=None):
prompt_token_ids = prompt_token_ids or []
return Sequence(
seq_id=0,
inputs=token_inputs(prompt_token_ids),
block_size=16,
)
def create_dummy_logprobs(
complete_sequence_token_ids: list[int]) -> list[dict[int, Logprob]]:
return [{
token_id: Logprob(logprob=0.0),
token_id + 1: Logprob(logprob=0.1)
} for token_id in complete_sequence_token_ids]
def create_dummy_prompt_logprobs(
complete_sequence_token_ids: list[int]
) -> list[Optional[dict[int, Any]]]:
# logprob for the first prompt token is None.
logprobs: list[Optional[dict[int, Any]]] = [None]
logprobs.extend(create_dummy_logprobs(complete_sequence_token_ids)[1:])
return logprobs
@pytest.mark.parametrize("complete_sequence", TRUTH)
@pytest.mark.parametrize("tokenizer_name", TOKENIZERS)
@pytest.mark.parametrize("skip_special_tokens", [True, False], indirect=True)
def test_decode_sequence_logprobs(complete_sequence: str,
complete_sequence_token_ids: list[int],
detokenizer: Detokenizer,
skip_special_tokens: bool):
"""Verify Detokenizer decodes logprobs correctly."""
sampling_params = SamplingParams(skip_special_tokens=skip_special_tokens,
logprobs=2)
# Run sequentially.
seq = create_sequence()
dummy_logprobs = create_dummy_logprobs(complete_sequence_token_ids)
sequential_logprobs_text_chosen_token: list[str] = []
sequential_logprobs_text_other_token: list[str] = []
for new_token, logprobs in zip(complete_sequence_token_ids,
dummy_logprobs):
seq.append_token_id(new_token, logprobs)
detokenizer.decode_sequence_inplace(seq, sampling_params)
sequential_logprobs_text_chosen_token.append(
seq.output_logprobs[-1][new_token].decoded_token)
sequential_logprobs_text_other_token.append(
seq.output_logprobs[-1][new_token + 1].decoded_token)
sequential_result = seq.output_text
assert sequential_result == "".join(sequential_logprobs_text_chosen_token)
assert sequential_result != "".join(sequential_logprobs_text_other_token)
if not skip_special_tokens:
# Text for logprobs for the chosen token should be the same as the
# generated text. Note that this will only be true if we skip
# special tokens.
assert sequential_result == complete_sequence
@pytest.mark.parametrize("complete_sequence", TRUTH)
@pytest.mark.parametrize("tokenizer_name", TOKENIZERS)
def test_decode_prompt_logprobs(complete_sequence: str,
complete_sequence_token_ids: list[int],
detokenizer: Detokenizer):
# We want to use skip_special_tokens=False here but Mistral tokenizers
# don't support that.
if complete_sequence not in SPECIAL_TOKS_TRUTH:
skip_special_tokens = True
elif not isinstance(detokenizer.tokenizer, MistralTokenizer):
skip_special_tokens = False
else:
pytest.skip("MistralTokenizers don't support "
"skip_special_tokens=False")
return
"""Verify Detokenizer decodes prompt logprobs correctly."""
sampling_params = SamplingParams(skip_special_tokens=skip_special_tokens,
prompt_logprobs=1)
# Run sequentially.
seq = create_sequence(complete_sequence_token_ids)
seq_group = SequenceGroup(request_id="1",
seqs=[seq],
sampling_params=sampling_params,
arrival_time=0.0)
dummy_logprobs = create_dummy_prompt_logprobs(complete_sequence_token_ids)
detokenizer.decode_prompt_logprobs_inplace(seq_group,
dummy_logprobs,
position_offset=0)
# First logprob is None.
decoded_prompt_logprobs: list[dict[int, Any]] = dummy_logprobs[
1:] # type: ignore
# decoded_prompt_logprobs doesn't contain the first token.
token_ids = complete_sequence_token_ids
tokenizer = detokenizer.tokenizer
text_full = tokenizer.decode(token_ids,
skip_special_tokens=skip_special_tokens)
text_first = tokenizer.decode(token_ids[0],
skip_special_tokens=skip_special_tokens)
text = text_full[len(text_first):]
# Text for logprobs for the chosen token should be the same as the
# prompt text. Note that the first logprob is None.
assert text == "".join([
logprobs[token_id].decoded_token
for token_id, logprobs in zip(token_ids[1:], decoded_prompt_logprobs)
])
assert text != "".join([
logprobs[token_id + 1].decoded_token
for token_id, logprobs in zip(token_ids[1:], decoded_prompt_logprobs)
])
+1 -1
View File
@@ -12,7 +12,7 @@ from partial_json_parser.core.options import Allow
from vllm.entrypoints.openai.protocol import (DeltaMessage, FunctionCall,
ToolCall)
from vllm.entrypoints.openai.tool_parsers import JambaToolParser
from vllm.transformers_utils.detokenizer import detokenize_incrementally
from vllm.transformers_utils.detokenizer_utils import detokenize_incrementally
from vllm.transformers_utils.tokenizer import AnyTokenizer, get_tokenizer
MODEL = "ai21labs/Jamba-tiny-dev"
@@ -13,7 +13,7 @@ from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
ToolCall)
from vllm.entrypoints.openai.tool_parsers.qwen3coder_tool_parser import (
Qwen3CoderToolParser)
from vllm.transformers_utils.detokenizer import detokenize_incrementally
from vllm.transformers_utils.detokenizer_utils import detokenize_incrementally
from vllm.transformers_utils.tokenizer import AnyTokenizer, get_tokenizer
MODEL = "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8"
+1 -1
View File
@@ -13,7 +13,7 @@ from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
DeltaMessage, FunctionCall,
ToolCall)
from vllm.entrypoints.openai.tool_parsers import SeedOssToolParser
from vllm.transformers_utils.detokenizer import detokenize_incrementally
from vllm.transformers_utils.detokenizer_utils import detokenize_incrementally
from vllm.transformers_utils.tokenizer import AnyTokenizer, get_tokenizer
# Use a common model that is likely to be available
+1 -1
View File
@@ -11,7 +11,7 @@ from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
DeltaMessage, FunctionCall,
ToolCall)
from vllm.entrypoints.openai.tool_parsers import xLAMToolParser
from vllm.transformers_utils.detokenizer import detokenize_incrementally
from vllm.transformers_utils.detokenizer_utils import detokenize_incrementally
from vllm.transformers_utils.tokenizer import AnyTokenizer, get_tokenizer
# Use a common model that is likely to be available
+1 -1
View File
@@ -12,9 +12,9 @@ from tests.v1.engine.utils import (NUM_PROMPT_LOGPROBS_UNDER_TEST,
STOP_STRINGS,
DummyOutputProcessorTestVectors,
MockEngineCore)
from vllm.logprobs import PromptLogprobs, SampleLogprobs
from vllm.outputs import CompletionOutput, RequestOutput
from vllm.sampling_params import RequestOutputKind, SamplingParams
from vllm.sequence import PromptLogprobs, SampleLogprobs
from vllm.transformers_utils.tokenizer import AnyTokenizer
from vllm.v1.engine import EngineCoreRequest
from vllm.v1.engine.output_processor import (OutputProcessor,