biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Chore] Remove redundant RequestPrompt (#30612)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-12-14 17:22:37 +08:00
committed by GitHub
parent f569c654e1
commit dcb31196da
12 changed files with 188 additions and 253 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
vllm/entrypoints/renderer.py
View File

@@ -12,9 +12,7 @@ import torch
from pydantic import Field
from vllm.config import ModelConfig
from vllm.inputs.data import EmbedsPrompt as EngineEmbedsPrompt
from vllm.inputs.data import TextPrompt as EngineTextPrompt
from vllm.inputs.data import TokensPrompt as EngineTokensPrompt
from vllm.inputs.data import EmbedsPrompt, TextPrompt, TokensPrompt
from vllm.inputs.parse import get_prompt_components, parse_raw_prompts
from vllm.tokenizers import TokenizerLike
from vllm.utils.async_utils import AsyncMicrobatchTokenizer
@@ -97,7 +95,7 @@ class BaseRenderer(ABC):
*,
prompt_or_prompts: str | list[str] | list[int] | list[list[int]],
config: RenderConfig,
) -> list[EngineTokensPrompt]:
) -> list[TokensPrompt]:
"""
Convert text or token inputs into engine-ready TokensPrompt objects.
@@ -115,7 +113,7 @@ class BaseRenderer(ABC):
(e.g., tokenization and length handling).
Returns:
list[EngineTokensPrompt]: Engine-ready token prompts.
list[TokensPrompt]: Engine-ready token prompts.
Raises:
ValueError: If input formats are invalid or length limits exceeded.
@@ -129,7 +127,7 @@ class BaseRenderer(ABC):
prompt_or_prompts: str | list[str] | list[int] | list[list[int]] | None = None,
prompt_embeds: bytes | list[bytes] | None = None,
config: RenderConfig,
) -> list[EngineTokensPrompt | EngineEmbedsPrompt]:
) -> list[TokensPrompt | EmbedsPrompt]:
"""
Convert text/token and/or base64-encoded embeddings inputs into
engine-ready prompt objects using a unified RenderConfig.
@@ -146,7 +144,7 @@ class BaseRenderer(ABC):
(e.g., tokenization and length handling).
Returns:
list[Union[EngineTokensPrompt, EngineEmbedsPrompt]]:
list[Union[TokensPrompt, EmbedsPrompt]]:
Engine-ready prompt objects.
Raises:
@@ -161,14 +159,14 @@ class BaseRenderer(ABC):
prompt_embeds: bytes | list[bytes],
truncate_prompt_tokens: Annotated[int, Field(ge=0)] | None = None,
cache_salt: str | None = None,
) -> list[EngineEmbedsPrompt]:
) -> list[EmbedsPrompt]:
"""Load and validate base64-encoded embeddings into prompt objects."""
if not self.model_config.enable_prompt_embeds:
raise ValueError(
"You must set `--enable-prompt-embeds` to input `prompt_embeds`."
)
def _load_and_validate_embed(embed: bytes) -> EngineEmbedsPrompt:
def _load_and_validate_embed(embed: bytes) -> EmbedsPrompt:
tensor = torch.load(
io.BytesIO(pybase64.b64decode(embed, validate=True)),
weights_only=True,
@@ -185,7 +183,7 @@ class BaseRenderer(ABC):
assert tensor.dim() == 2
if truncate_prompt_tokens is not None:
tensor = tensor[-truncate_prompt_tokens:]
embeds_prompt = EngineEmbedsPrompt(prompt_embeds=tensor)
embeds_prompt = EmbedsPrompt(prompt_embeds=tensor)
if cache_salt is not None:
embeds_prompt["cache_salt"] = cache_salt
return embeds_prompt
@@ -213,7 +211,7 @@ class CompletionRenderer(BaseRenderer):
*,
prompt_or_prompts: str | list[str] | list[int] | list[list[int]],
config: RenderConfig,
) -> list[EngineTokensPrompt]:
) -> list[TokensPrompt]:
"""Implementation of prompt rendering for completion-style requests.
Uses async tokenizer pooling for improved performance. See base class
@@ -240,7 +238,7 @@ class CompletionRenderer(BaseRenderer):
prompt_or_prompts: str | list[str] | list[int] | list[list[int]] | None = None,
prompt_embeds: bytes | list[bytes] | None = None,
config: RenderConfig,
) -> list[EngineTokensPrompt | EngineEmbedsPrompt]:
) -> list[TokensPrompt | EmbedsPrompt]:
"""
Render text/token prompts and/or precomputed embedding prompts. At
least one of `prompt_or_prompts` or `prompt_embeds` must be provided.
@@ -249,7 +247,7 @@ class CompletionRenderer(BaseRenderer):
if truncate_prompt_tokens == 0:
return []
rendered: list[EngineTokensPrompt | EngineEmbedsPrompt] = []
rendered: list[TokensPrompt | EmbedsPrompt] = []
if prompt_embeds is not None:
rendered.extend(
@@ -281,10 +279,10 @@ class CompletionRenderer(BaseRenderer):
async def _create_prompt(
self,
prompt_input: EngineTextPrompt | EngineTokensPrompt,
prompt_input: TextPrompt | TokensPrompt,
config: RenderConfig,
truncate_prompt_tokens: int | None,
) -> EngineTokensPrompt:
) -> TokensPrompt:
prompt, prompt_token_ids, _ = get_prompt_components(prompt_input)
if prompt_token_ids is not None:
@@ -317,7 +315,7 @@ class CompletionRenderer(BaseRenderer):
truncate_prompt_tokens: int | None,
add_special_tokens: bool,
cache_salt: str | None,
) -> EngineTokensPrompt:
) -> TokensPrompt:
"""Tokenize text input asynchronously."""
async_tokenizer = self._get_async_tokenizer()
@@ -350,7 +348,7 @@ class CompletionRenderer(BaseRenderer):
truncate_prompt_tokens: int | None,
cache_salt: str | None,
needs_detokenization: bool | None = False,
) -> EngineTokensPrompt:
) -> TokensPrompt:
"""Optionally detokenize token IDs and build a tokens prompt."""
token_ids = self._maybe_apply_truncation(token_ids, truncate_prompt_tokens)
@@ -392,8 +390,8 @@ class CompletionRenderer(BaseRenderer):
max_length: int | None = None,
cache_salt: str | None = None,
prompt: str | None = None,
) -> EngineTokensPrompt:
"""Create validated EngineTokensPrompt."""
) -> TokensPrompt:
"""Create validated TokensPrompt."""
if max_length is not None and len(token_ids) > max_length:
raise ValueError(
f"This model's maximum context length is {max_length} tokens. "
@@ -401,7 +399,7 @@ class CompletionRenderer(BaseRenderer):
"Please reduce the length of the input messages."
)
tokens_prompt = EngineTokensPrompt(prompt_token_ids=token_ids)
tokens_prompt = TokensPrompt(prompt_token_ids=token_ids)
if cache_salt is not None:
tokens_prompt["cache_salt"] = cache_salt
if prompt is not None: