[Bugfix] Treat max_tokens as output upper bound, not input-space reservation (#57035)

Signed-off-by: YuvrajSinghBhadoria2 <245782664+YuvrajSinghBhadoria2@users.noreply.github.com>
Signed-off-by: YuvrajSinghBhadoria2 <yuvrajsinghbhado2030@gmail.com>
Signed-off-by: Yuvraj Singh Bhadoria <yuvrajsinghbhado2030@gmail.com>
Co-authored-by: YuvrajSinghBhadoria2 <245782664+YuvrajSinghBhadoria2@users.noreply.github.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Yuvraj Singh Bhadoria
2026-10-10 01:15:22 +08:00
committed by GitHub
co-authored by YuvrajSinghBhadoria2 Cyrus Leung Claude Opus 4.8
parent 89ac8d971b
commit 1a3a7a8f01
7 changed files with 252 additions and 32 deletions
+3 -3
View File
@@ -183,15 +183,15 @@ class EmbeddingTokenizeParamsMixin(PoolingTokenizeParamsMixin):
if pooler_config.enable_chunked_processing:
max_total_tokens = pooler_config.max_embed_len
else:
max_embed_len = pooler_config.max_embed_len or default_max_total_tokens
max_output_tokens = default_max_total_tokens - max_embed_len
max_total_tokens = (
pooler_config.max_embed_len or default_max_total_tokens
)
tok_params = self._build_pooling_tok_params(
model_config,
add_special_tokens=self.add_special_tokens,
max_total_tokens=max_total_tokens,
max_output_tokens=max_output_tokens,
max_output_tokens_param="max_model_len - max_embed_len",
)
if (
+49 -22
View File
@@ -207,11 +207,38 @@ class TokenizeParams:
@property
def max_input_tokens(self) -> int | None:
"""Maximum allowed number of input tokens."""
"""Maximum allowed number of input tokens.
``max_output_tokens`` is *not* subtracted here: the value represents the
actual model context length, not a reservation of output space.
Downstream clamping of the sampling ``max_tokens`` is handled by
``get_max_tokens()`` in the entrypoint layer.
"""
if self.max_total_tokens is None:
return None
return self.max_total_tokens - self.max_output_tokens
return self.max_total_tokens
@property
def max_truncation_tokens(self) -> int | None:
"""Maximum prompt length that truncation may reduce a prompt to.
For generative requests this is `max_input_tokens` less one, holding a
token of context back for output. Truncating to the full
`max_total_tokens` would leave `get_max_tokens()` clamping the sampling
budget to 0, turning an over-long prompt into a silent zero-token
generation.
Pooling requests set `max_output_tokens = 0` and emit an embedding
rather than generated tokens, so they keep the full input budget.
"""
max_input_tokens = self.max_input_tokens
if max_input_tokens is None:
return None
if self.max_output_tokens == 0:
return max_input_tokens
return max(0, max_input_tokens - 1)
def __post_init__(self) -> None:
max_total_tokens = self.max_total_tokens
@@ -247,14 +274,15 @@ class TokenizeParams:
):
raise VLLMValidationError(
f"{self.truncate_prompt_tokens_param}={truncate_prompt_tokens} "
f"cannot be greater than {self.max_total_tokens_param} - "
f"{self.max_output_tokens_param} = {max_input_tokens}. "
f"cannot be greater than {self.max_total_tokens_param} "
f"= {max_input_tokens}. "
f"Please request a smaller truncation size.",
parameter=self.truncate_prompt_tokens_param,
value=truncate_prompt_tokens,
)
def with_kwargs(self, **tokenization_kwargs: Any):
has_max_length = "max_length" in tokenization_kwargs
max_length = tokenization_kwargs.pop("max_length", self.max_input_tokens)
pad_prompt_tokens = tokenization_kwargs.pop(
"pad_prompt_tokens", self.pad_prompt_tokens
@@ -301,13 +329,17 @@ class TokenizeParams:
max_total_tokens = self.max_total_tokens
if has_max_length and max_total_tokens is not None and max_length is not None:
max_output_tokens = max_total_tokens - max_length
else:
# No explicit ``max_length`` override: preserve the original output
# budget instead of recomputing it from ``max_input_tokens`` (which
# is now the full context length).
max_output_tokens = self.max_output_tokens
return TokenizeParams(
max_total_tokens=max_total_tokens,
max_output_tokens=(
0
if max_total_tokens is None or max_length is None
else max_total_tokens - max_length
),
max_output_tokens=max_output_tokens,
pad_prompt_tokens=pad_prompt_tokens,
truncate_prompt_tokens=truncate_prompt_tokens,
truncation_side=truncation_side,
@@ -320,7 +352,7 @@ class TokenizeParams:
"""The arguments to pass to `tokenizer.encode`."""
max_length = self.truncate_prompt_tokens
if max_length is not None and max_length < 0:
max_length = self.max_input_tokens
max_length = self.max_truncation_tokens
elif max_length is None and self.max_input_tokens is not None:
# This prevents tokenization from taking up more resources than necessary
# while still failing `self._token_len_check` as expected by users
@@ -355,13 +387,11 @@ class TokenizeParams:
if len(text) > max_input_chars:
raise VLLMValidationError(
f"This model's maximum context length is "
f"{self.max_total_tokens} tokens. However, you requested "
f"{self.max_output_tokens} output tokens and your prompt "
f"{self.max_total_tokens} tokens. However, your prompt "
f"contains {len(text)} characters (more than "
f"{max_input_chars} characters, which is the upper bound "
f"for {max_input_tokens} input tokens). "
f"Please reduce the length of the input prompt or the "
f"number of requested output tokens.",
f"Please reduce the length of the input prompt.",
parameter="input_text",
value=len(text),
)
@@ -459,7 +489,7 @@ class TokenizeParams:
"""
max_length = self.truncate_prompt_tokens
if max_length is not None and max_length < 0:
max_length = self.max_input_tokens
max_length = self.max_truncation_tokens
if max_length is None or max_length >= length:
return None
@@ -494,15 +524,12 @@ class TokenizeParams:
# max_input_tokens + 1 (see get_encode_kwargs), so the
# actual prompt length could be larger.
qualifier = "at least " if token_count == max_input_tokens + 1 else ""
total = token_count + self.max_output_tokens
raise VLLMValidationError(
f"This model's maximum context length is "
f"{self.max_total_tokens} tokens. However, you requested "
f"{self.max_output_tokens} output tokens and your prompt "
f"contains {qualifier}{token_count} input tokens, "
f"for a total of {qualifier}{total} tokens. "
f"Please reduce the length of the input prompt or the "
f"number of requested output tokens.",
f"{self.max_total_tokens} tokens. However, your prompt "
f"contains {qualifier}{token_count} input tokens, which "
f"exceeds this limit even without any output tokens. "
f"Please reduce the length of the input prompt.",
parameter="input_tokens",
value=token_count,
)