d636456b1f · tduncan/Nvidia-Optimization-Stack-VLLM

benchmarks/backend_request_func.py

-4

@@ -630,10 +630,6 @@ def get_tokenizer(

630 pretrained_model_name_or_path

631 ):

632 pretrained_model_name_or_path = get_model(pretrained_model_name_or_path)

633 - if tokenizer_mode == "slow":

634 - if kwargs.get("use_fast", False):

635 - raise ValueError("Cannot use the fast tokenizer in slow tokenizer mode.")

636 - kwargs["use_fast"] = False

633 if tokenizer_mode == "mistral":

634 try:

635 from vllm.tokenizers.mistral import MistralTokenizer

examples/generate/multimodal/vision_language_multi_image_offline.py

-1

@@ -55,7 +55,6 @@ def load_aria(question: str, image_urls: list[str]) -> ModelRequestData:

55 model_name = "rhymes-ai/Aria"

56 engine_args = EngineArgs(

57 model=model_name,

58 - tokenizer_mode="slow",

58 trust_remote_code=True,

59 dtype="bfloat16",

60 limit_mm_per_prompt={"image": len(image_urls)},

rust/src/bench/src/cli.rs

+1 -1

@@ -176,7 +176,7 @@ pub struct BenchServeArgs {

176 #[arg(long)]

177 pub tokenizer: Option<String>,

178

179 - /// Tokenizer mode (auto, hf, slow, mistral). Accepted for Python CLI

179 + /// Tokenizer mode (auto, hf, mistral). Accepted for Python CLI

180 /// compatibility; non-auto values are ignored with a warning.

181 #[arg(long, default_value = "auto")]

182 pub tokenizer_mode: String,

tests/renderers/test_token_offsets.py

+1 -1

@@ -20,7 +20,7 @@ def fast_tokenizer():

20 """gpt2 ships a Fast tokenizer; use it to test the offsets happy path."""

21 from transformers import AutoTokenizer

22

23 - return AutoTokenizer.from_pretrained("openai-community/gpt2", use_fast=True)

23 + return AutoTokenizer.from_pretrained("openai-community/gpt2")

24

25

26 def _make_base_renderer_with(tokenizer):

tests/tokenizers_/test_basic.py

+1 -1

@@ -23,7 +23,7 @@ def _assert_tokenizer_like(tokenizer: object):

23

24

25 def test_tokenizer_like_protocol():

26 - tokenizer = get_tokenizer("openai-community/gpt2", use_fast=True)

26 + tokenizer = get_tokenizer("openai-community/gpt2")

27 assert isinstance(tokenizer, TokenizersBackend)

28 _assert_tokenizer_like(tokenizer)

29

tests/tokenizers_/test_registry.py

+5 -16

@@ -69,21 +69,11 @@ def test_resolve_tokenizer_args_idempotent(runner_type):

69 )

70

71

72 -@pytest.mark.parametrize(

73 - ("tokenizer_mode", "input_kwargs"),

74 - [

75 - ("hf", {}),

76 - ("hf", {"mistral_format": False}),

77 - ("slow", {}),

78 - ("slow", {"mistral_format": False}),

79 - ],

80 -)

81 -def test_resolve_tokenizer_args_forces_hf_mistral_format_false(

82 - tokenizer_mode, input_kwargs

83 -):

72 +@pytest.mark.parametrize("input_kwargs", [{}, {"mistral_format": False}])

73 +def test_resolve_tokenizer_args_forces_hf_mistral_format_false(input_kwargs):

74 resolved_mode, _, _, kwargs = resolve_tokenizer_args(

75 "mistralai/Mistral-Nemo-Instruct-2407",

86 - tokenizer_mode=tokenizer_mode,

76 + tokenizer_mode="hf",

77 **input_kwargs,

78 )

79

@@ -91,15 +81,14 @@ def test_resolve_tokenizer_args_forces_hf_mistral_format_false(

81 assert kwargs["mistral_format"] is False

82

83

94 -@pytest.mark.parametrize("tokenizer_mode", ["hf", "slow"])

95 -def test_resolve_tokenizer_args_rejects_hf_mistral_format_true(tokenizer_mode):

84 +def test_resolve_tokenizer_args_rejects_hf_mistral_format_true():

85 with pytest.raises(

86 ValueError,

87 match="mistral_format=True is not supported with tokenizer_mode='hf'",

88 ):

89 resolve_tokenizer_args(

90 "mistralai/Mistral-Nemo-Instruct-2407",

102 - tokenizer_mode=tokenizer_mode,

91 + tokenizer_mode="hf",

92 mistral_format=True,

93 )

94

vllm/benchmarks/serve.py

-1

@@ -1695,7 +1695,6 @@ def add_cli_args(parser: FlexibleArgumentParser):

1695 - "auto" will use the tokenizer from `mistral_common` for Mistral models

1696 if available, otherwise it will use the "hf" tokenizer.\n

1697 - "hf" will use the fast tokenizer if available.\n

1698 - - "slow" will always use the slow tokenizer.\n

1698 - "mistral" will always use the tokenizer from `mistral_common`.\n

1699 - "deepseek_v32" will always use the tokenizer from `deepseek_v32`.\n

1700 - Other custom values can be supported via plugins.""",

vllm/config/model.py

-2

@@ -89,7 +89,6 @@ ConvertOption = Literal["auto", ConvertType]

89 TokenizerMode = Literal[

90 "auto",

91 "hf",

92 - "slow",

92 "mistral",

93 "deepseek_v32",

94 "deepseek_v4",

@@ -148,7 +147,6 @@ class ModelConfig:

147 - "auto" will use the tokenizer from `mistral_common` for Mistral models

148 if available, otherwise it will use the "hf" tokenizer.

149 - "hf" will use the fast tokenizer if available.

151 - - "slow" will always use the slow tokenizer.

150 - "mistral" will always use the tokenizer from `mistral_common`.

151 - "deepseek_v32" will always use the tokenizer from `deepseek_v32`.

152 - "deepseek_v4" will always use the tokenizer from `deepseek_v4`.

vllm/entrypoints/llm.py

+2 -2

@@ -76,8 +76,8 @@ class LLM(BeamSearchOfflineMixin, PoolingOfflineMixin, OfflineInferenceMixin):

76 Args:

77 model: The name or path of a HuggingFace Transformers model.

78 tokenizer: The name or path of a HuggingFace Transformers tokenizer.

79 - tokenizer_mode: The tokenizer mode. "auto" will use the fast tokenizer

80 - if available, and "slow" will always use the slow tokenizer.

79 + tokenizer_mode: The tokenizer mode. See

80 + [ModelConfig.tokenizer_mode][vllm.config.ModelConfig.tokenizer_mode].

81 skip_tokenizer_init: If true, skip initialization of tokenizer and

82 detokenizer. Expect valid prompt_token_ids and None for prompt

83 from the input.

vllm/renderers/hf.py

+1 -2

@@ -966,8 +966,7 @@ class HfRenderer(BaseRenderer[HfTokenizer]):

966 )

967

968 def _can_produce_offsets(self) -> bool:

969 - # HF tokenizers may be slow (use_fast=False); only fast tokenizers

970 - # expose offset_mapping.

969 + # Only fast tokenizers expose offset_mapping.

970 return self.tokenizer is not None and self.tokenizer.is_fast

971

972 def render_messages(

vllm/tokenizers/registry.py

-7

@@ -145,13 +145,6 @@ def resolve_tokenizer_args(

145 else:

146 assert_never(runner_type)

147

148 - if tokenizer_mode == "slow":

149 - if kwargs.get("use_fast", False):

150 - raise ValueError("Cannot use the fast tokenizer in slow tokenizer mode.")

151 -

152 - tokenizer_mode = "hf"

153 - kwargs["use_fast"] = False

154 -

148 # Try to use official Mistral tokenizer if possible

149 if (

150 tokenizer_mode == "auto"

vllm/transformers_utils/processors/moondream3.py

-1

@@ -209,7 +209,6 @@ class Moondream3Processor(ProcessorMixin):

209 "revision",

210 "subfolder",

211 "token",

212 - "use_fast",

212 ):

213 if key in kwargs:

214 tokenizer_kwargs[key] = kwargs[key]

[Frontend] Remove the slow tokenizer mode (#58545)