a358e4dffe · tduncan/Nvidia-Optimization-Stack-VLLM

vllm/engine/protocol.py

+2 -2

@@ -11,7 +11,7 @@ from vllm.lora.request import LoRARequest

11 from vllm.outputs import PoolingRequestOutput, RequestOutput

12 from vllm.plugins.io_processors import IOProcessor

13 from vllm.pooling_params import PoolingParams

14 -from vllm.renderers import RendererLike

14 +from vllm.renderers import BaseRenderer

15 from vllm.sampling_params import SamplingParams

16 from vllm.tasks import SupportedTask

17 from vllm.v1.engine import EngineCoreRequest

@@ -28,7 +28,7 @@ class EngineClient(ABC):

28

29 @property

30 @abstractmethod

31 - def renderer(self) -> RendererLike: ...

31 + def renderer(self) -> BaseRenderer: ...

32

33 @property

34 @abstractmethod

vllm/renderers/__init__.py

+2 -2

@@ -2,11 +2,11 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 from .params import ChatParams, TokenizeParams, merge_kwargs

5 -from .protocol import RendererLike

5 +from .protocol import BaseRenderer

6 from .registry import RendererRegistry, renderer_from_config

7

8 __all__ = [

9 - "RendererLike",

9 + "BaseRenderer",

10 "RendererRegistry",

11 "renderer_from_config",

12 "ChatParams",

vllm/renderers/deepseek_v32.py

+4 -6

@@ -15,18 +15,18 @@ from vllm.tokenizers import cached_get_tokenizer

15 from vllm.tokenizers.deepseek_v32 import DeepseekV32Tokenizer

16

17 from .params import ChatParams

18 -from .protocol import RendererLike

18 +from .protocol import BaseRenderer

19

20 logger = init_logger(__name__)

21

22

23 -class DeepseekV32Renderer(RendererLike):

23 +class DeepseekV32Renderer(BaseRenderer):

24 @classmethod

25 def from_config(

26 cls,

27 config: ModelConfig,

28 tokenizer_kwargs: dict[str, Any],

29 - ) -> "RendererLike":

29 + ) -> "BaseRenderer":

30 return cls(config, tokenizer_kwargs)

31

32 def __init__(

@@ -34,9 +34,7 @@ class DeepseekV32Renderer(RendererLike):

34 config: ModelConfig,

35 tokenizer_kwargs: dict[str, Any],

36 ) -> None:

37 - super().__init__()

38 -

39 - self.config = config

37 + super().__init__(config)

38

39 if config.skip_tokenizer_init:

40 tokenizer = None

vllm/renderers/grok2.py

+4 -6

@@ -15,18 +15,18 @@ from vllm.tokenizers import cached_get_tokenizer

15 from vllm.tokenizers.grok2 import Grok2Tokenizer

16

17 from .params import ChatParams

18 -from .protocol import RendererLike

18 +from .protocol import BaseRenderer

19

20 logger = init_logger(__name__)

21

22

23 -class Grok2Renderer(RendererLike):

23 +class Grok2Renderer(BaseRenderer):

24 @classmethod

25 def from_config(

26 cls,

27 config: ModelConfig,

28 tokenizer_kwargs: dict[str, Any],

29 - ) -> "RendererLike":

29 + ) -> "BaseRenderer":

30 return cls(config, tokenizer_kwargs)

31

32 def __init__(

@@ -34,9 +34,7 @@ class Grok2Renderer(RendererLike):

34 config: ModelConfig,

35 tokenizer_kwargs: dict[str, Any],

36 ) -> None:

37 - super().__init__()

38 -

39 - self.config = config

37 + super().__init__(config)

38

39 if config.skip_tokenizer_init:

40 tokenizer = None

vllm/renderers/hf.py

+4 -5

@@ -34,7 +34,7 @@ from vllm.transformers_utils.processor import cached_get_processor

34 from vllm.utils.func_utils import supports_kw

35

36 from .params import ChatParams

37 -from .protocol import RendererLike

37 +from .protocol import BaseRenderer

38

39 if TYPE_CHECKING:

40 from vllm.multimodal.inputs import MultiModalDataDict, MultiModalUUIDDict

@@ -584,13 +584,13 @@ def replace_vision_chunk_video_placeholder(

584 return prompt_raw

585

586

587 -class HfRenderer(RendererLike):

587 +class HfRenderer(BaseRenderer):

588 @classmethod

589 def from_config(

590 cls,

591 config: ModelConfig,

592 tokenizer_kwargs: dict[str, Any],

593 - ) -> "RendererLike":

593 + ) -> "BaseRenderer":

594 return cls(config, tokenizer_kwargs)

595

596 def __init__(

@@ -598,9 +598,8 @@ class HfRenderer(RendererLike):

598 config: ModelConfig,

599 tokenizer_kwargs: dict[str, Any],

600 ) -> None:

601 - super().__init__()

601 + super().__init__(config)

602

603 - self.config = config

603 self.use_unified_vision_chunk = getattr(

604 config.hf_config, "use_unified_vision_chunk", False

605 )

vllm/renderers/mistral.py

+4 -6

@@ -17,7 +17,7 @@ from vllm.tokenizers.mistral import MistralTokenizer

17 from vllm.utils.async_utils import make_async

18

19 from .params import ChatParams

20 -from .protocol import RendererLike

20 +from .protocol import BaseRenderer

21

22 logger = init_logger(__name__)

23

@@ -49,13 +49,13 @@ def safe_apply_chat_template(

49 raise ValueError(str(e)) from e

50

51

52 -class MistralRenderer(RendererLike):

52 +class MistralRenderer(BaseRenderer):

53 @classmethod

54 def from_config(

55 cls,

56 config: ModelConfig,

57 tokenizer_kwargs: dict[str, Any],

58 - ) -> "RendererLike":

58 + ) -> "BaseRenderer":

59 return cls(config, tokenizer_kwargs)

60

61 def __init__(

@@ -63,9 +63,7 @@ class MistralRenderer(RendererLike):

63 config: ModelConfig,

64 tokenizer_kwargs: dict[str, Any],

65 ) -> None:

66 - super().__init__()

67 -

68 - self.config = config

66 + super().__init__(config)

67

68 if config.skip_tokenizer_init:

69 tokenizer = None

vllm/renderers/protocol.py

+16 -8

@@ -1,7 +1,8 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 import asyncio

4 -from typing import TYPE_CHECKING, Any, Protocol

4 +from abc import ABC, abstractmethod

5 +from typing import TYPE_CHECKING, Any

6

7 from vllm.inputs import EmbedsPrompt, TextPrompt, TokensPrompt

8 from vllm.tokenizers import TokenizerLike

@@ -19,19 +20,26 @@ if TYPE_CHECKING:

20 )

21

22

22 -class RendererLike(Protocol):

23 - config: "ModelConfig"

24 - _async_tokenizer: AsyncMicrobatchTokenizer

25 -

23 +class BaseRenderer(ABC):

24 @classmethod

25 + @abstractmethod

26 def from_config(

27 cls,

28 config: "ModelConfig",

29 tokenizer_kwargs: dict[str, Any],

31 - ) -> "RendererLike":

30 + ) -> "BaseRenderer":

31 raise NotImplementedError

32

33 + def __init__(self, config: "ModelConfig") -> None:

34 + super().__init__()

35 +

36 + self.config = config

37 +

38 + # Lazy initialization since offline LLM doesn't use async

39 + self._async_tokenizer: AsyncMicrobatchTokenizer | None = None

40 +

41 @property

42 + @abstractmethod

43 def tokenizer(self) -> TokenizerLike | None:

44 raise NotImplementedError

45

@@ -43,8 +51,7 @@ class RendererLike(Protocol):

51 return tokenizer

52

53 def get_async_tokenizer(self) -> AsyncMicrobatchTokenizer:

46 - # Lazy initialization since offline LLM doesn't use async

47 - if not hasattr(self, "_async_tokenizer"):

54 + if self._async_tokenizer is None:

55 self._async_tokenizer = AsyncMicrobatchTokenizer(self.get_tokenizer())

56

57 return self._async_tokenizer

@@ -104,6 +111,7 @@ class RendererLike(Protocol):

111 ) -> list[TextPrompt | TokensPrompt | EmbedsPrompt]:

112 return self.render_completions(prompt_input, prompt_embeds)

113

114 + @abstractmethod

115 def render_messages(

116 self,

117 messages: list["ChatCompletionMessageParam"],

vllm/renderers/registry.py

+3 -3

@@ -7,7 +7,7 @@ from vllm.logger import init_logger

7 from vllm.tokenizers.registry import tokenizer_args_from_config

8 from vllm.utils.import_utils import resolve_obj_by_qualname

9

10 -from .protocol import RendererLike

10 +from .protocol import BaseRenderer

11

12 if TYPE_CHECKING:

13 from vllm.config import ModelConfig

@@ -43,7 +43,7 @@ class RendererRegistry:

43

44 return None

45

46 - def load_renderer_cls(self, renderer_mode: str) -> type[RendererLike]:

46 + def load_renderer_cls(self, renderer_mode: str) -> type[BaseRenderer]:

47 if renderer_mode not in self.renderers:

48 raise ValueError(f"No renderer registered for {renderer_mode=!r}.")

49

@@ -57,7 +57,7 @@ class RendererRegistry:

57 renderer_mode: str,

58 config: "ModelConfig",

59 tokenizer_kwargs: dict[str, Any],

60 - ) -> RendererLike:

60 + ) -> BaseRenderer:

61 renderer_cls = self.load_renderer_cls(renderer_mode)

62 return renderer_cls.from_config(config, tokenizer_kwargs)

63

vllm/renderers/terratorch.py

+4 -6

@@ -14,24 +14,22 @@ from vllm.logger import init_logger

14 from vllm.tokenizers import TokenizerLike

15

16 from .params import ChatParams

17 -from .protocol import RendererLike

17 +from .protocol import BaseRenderer

18

19 logger = init_logger(__name__)

20

21

22 -class TerratorchRenderer(RendererLike):

22 +class TerratorchRenderer(BaseRenderer):

23 @classmethod

24 def from_config(

25 cls,

26 config: "ModelConfig",

27 tokenizer_kwargs: dict[str, Any],

28 - ) -> "RendererLike":

28 + ) -> "BaseRenderer":

29 return cls(config)

30

31 def __init__(self, config: ModelConfig) -> None:

32 - super().__init__()

33 -

34 - self.config = config

32 + super().__init__(config)

33

34 if not config.skip_tokenizer_init:

35 raise ValueError("Terratorch renderer requires `skip_tokenizer_init=True`")

vllm/v1/engine/async_llm.py

+2 -2

@@ -24,7 +24,7 @@ from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry

24 from vllm.outputs import STREAM_FINISHED, PoolingRequestOutput, RequestOutput

25 from vllm.plugins.io_processors import get_io_processor

26 from vllm.pooling_params import PoolingParams

27 -from vllm.renderers import RendererLike, merge_kwargs

27 +from vllm.renderers import BaseRenderer, merge_kwargs

28 from vllm.sampling_params import RequestOutputKind, SamplingParams

29 from vllm.tasks import SupportedTask

30 from vllm.tokenizers import TokenizerLike

@@ -844,7 +844,7 @@ class AsyncLLM(EngineClient):

844 return self.input_processor.get_tokenizer()

845

846 @property

847 - def renderer(self) -> RendererLike:

847 + def renderer(self) -> BaseRenderer:

848 return self.input_processor.renderer

849

850 async def is_tracing_enabled(self) -> bool:

vllm/v1/engine/input_processor.py

+2 -2

@@ -29,7 +29,7 @@ from vllm.multimodal.parse import ModalityDataItems, MultiModalDataItems

29 from vllm.multimodal.processing.context import set_request_id

30 from vllm.multimodal.utils import argsort_mm_positions

31 from vllm.pooling_params import PoolingParams

32 -from vllm.renderers import RendererLike

32 +from vllm.renderers import BaseRenderer

33 from vllm.sampling_params import _SAMPLING_EPS, SamplingParams

34 from vllm.tokenizers import TokenizerLike

35 from vllm.tokenizers.mistral import MistralTokenizer

@@ -96,7 +96,7 @@ class InputProcessor:

96 return self.input_preprocessor.get_tokenizer()

97

98 @property

99 - def renderer(self) -> RendererLike:

99 + def renderer(self) -> BaseRenderer:

100 return self.input_preprocessor.renderer

101

102 def _validate_logprobs(

vllm/v1/engine/llm_engine.py

+2 -2

@@ -21,7 +21,7 @@ from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry

21 from vllm.outputs import PoolingRequestOutput, RequestOutput

22 from vllm.plugins.io_processors import get_io_processor

23 from vllm.pooling_params import PoolingParams

24 -from vllm.renderers import RendererLike

24 +from vllm.renderers import BaseRenderer

25 from vllm.sampling_params import SamplingParams

26 from vllm.tasks import SupportedTask

27 from vllm.tokenizers import TokenizerLike

@@ -367,7 +367,7 @@ class LLMEngine:

367 return self.input_processor.get_tokenizer()

368

369 @property

370 - def renderer(self) -> RendererLike:

370 + def renderer(self) -> BaseRenderer:

371 return self.input_processor.renderer

372

373 def do_log_stats(self) -> None:

[Refactor] Make Renderer an abstract class (#33479)