[Refactor] Pass Renderer to Input Processor (#34329)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-02-11 19:38:11 -08:00
committed by GitHub
parent ced2a92f40
commit b96f7314b4
20 changed files with 107 additions and 106 deletions
+16 -15
View File
@@ -242,11 +242,10 @@ class OpenAIServing:
self.log_error_stack = log_error_stack
self.input_processor = self.models.input_processor
self.io_processor = self.models.io_processor
self.renderer = self.models.renderer
self.model_config = self.models.model_config
self.max_model_len = self.model_config.max_model_len
self.model_config = engine_client.model_config
self.renderer = engine_client.renderer
self.io_processor = engine_client.io_processor
self.input_processor = engine_client.input_processor
async def beam_search(
self,
@@ -537,7 +536,7 @@ class OpenAIServing:
if (
truncate_prompt_tokens is not None
and truncate_prompt_tokens > self.max_model_len
and truncate_prompt_tokens > self.model_config.max_model_len
):
return self.create_error_response(
"truncate_prompt_tokens value is "
@@ -844,6 +843,7 @@ class OpenAIServing:
input_text: str,
) -> TokensPrompt:
token_num = len(input_ids)
max_model_len = self.model_config.max_model_len
# Note: EmbeddingRequest, ClassificationRequest,
# and ScoreRequest doesn't have max_tokens
@@ -862,7 +862,7 @@ class OpenAIServing:
):
# Note: input length can be up to the entire model context length
# since these requests don't generate tokens.
if token_num > self.max_model_len:
if token_num > max_model_len:
operations: dict[type[AnyRequest], str] = {
ScoreDataRequest: "score",
ScoreTextRequest: "score",
@@ -873,7 +873,7 @@ class OpenAIServing:
operation = operations.get(type(request), "embedding generation")
raise VLLMValidationError(
f"This model's maximum context length is "
f"{self.max_model_len} tokens. However, you requested "
f"{max_model_len} tokens. However, you requested "
f"{token_num} tokens in the input for {operation}. "
f"Please reduce the length of the input.",
parameter="input_tokens",
@@ -898,22 +898,22 @@ class OpenAIServing:
# Note: input length can be up to model context length - 1 for
# completion-like requests.
if token_num >= self.max_model_len:
if token_num >= max_model_len:
raise VLLMValidationError(
f"This model's maximum context length is "
f"{self.max_model_len} tokens. However, your request has "
f"{max_model_len} tokens. However, your request has "
f"{token_num} input tokens. Please reduce the length of "
"the input messages.",
parameter="input_tokens",
value=token_num,
)
if max_tokens is not None and token_num + max_tokens > self.max_model_len:
if max_tokens is not None and token_num + max_tokens > max_model_len:
raise VLLMValidationError(
"'max_tokens' or 'max_completion_tokens' is too large: "
f"{max_tokens}. This model's maximum context length is "
f"{self.max_model_len} tokens and your request has "
f"{token_num} input tokens ({max_tokens} > {self.max_model_len}"
f"{max_model_len} tokens and your request has "
f"{token_num} input tokens ({max_tokens} > {max_model_len}"
f" - {token_num}).",
parameter="max_tokens",
value=max_tokens,
@@ -1089,6 +1089,7 @@ class OpenAIServing:
priority: int = 0,
trace_headers: Mapping[str, str] | None = None,
):
max_model_len = self.model_config.max_model_len
prompt_text = self._extract_prompt_text(engine_prompt)
orig_priority = priority
@@ -1148,7 +1149,7 @@ class OpenAIServing:
token_ids = context.render_for_completion()
engine_prompt = TokensPrompt(prompt_token_ids=token_ids)
sampling_params.max_tokens = self.max_model_len - len(token_ids)
sampling_params.max_tokens = max_model_len - len(token_ids)
elif isinstance(context, ParsableContext):
engine_prompts = await self._render_next_turn(
context.request,
@@ -1162,7 +1163,7 @@ class OpenAIServing:
prompt_text = self._extract_prompt_text(engine_prompt)
sampling_params.max_tokens = get_max_tokens(
self.max_model_len,
max_model_len,
context.request.max_output_tokens,
self._extract_prompt_len(engine_prompt),
self.default_sampling_params, # type: ignore