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

[Misc] Rename think_start_str/think_end_str to reasoning_start_str/reasoning_end_str (#38242)

Browse Source 
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
(cherry picked from commit cbe7d18096)
This commit is contained in:
Chauncey
2026-04-02 00:56:45 +08:00
committed by khluu
parent 29982d48b3
commit 3a30a1a6a8
5 changed files with 55 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
docs/features/reasoning_outputs.md
View File

@@ -244,12 +244,12 @@ response = client.chat.completions.create(
Some models, such as [Qwen3](https://qwen.readthedocs.io/en/latest/getting_started/quickstart.html#thinking-budget), [DeepSeek](https://www.alibabacloud.com/help/en/model-studio/deep-thinking), and [Nemotron3](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16), support a thinking budget that limits the maximum number of tokens used for reasoning.
Token counting starts from `think_start_str`. Once the reasoning token count reaches the configured `thinking_token_budget`, vLLM forces the model to produce `think_end_str`, effectively terminating the reasoning block.
Token counting starts from `reasoning_start_str`. Once the reasoning token count reaches the configured `thinking_token_budget`, vLLM forces the model to produce `reasoning_end_str`, effectively terminating the reasoning block.
To use this feature:
- `--reasoning-parser` enables reasoning extraction.
- `--reasoning-config` defines the reasoning boundary tokens (e.g., `think_start_str`, `think_end_str`).
- `--reasoning-config` defines the reasoning boundary tokens (e.g., `reasoning_start_str`, `reasoning_end_str`).
- `thinking_token_budget` (a sampling parameter) sets the per-request reasoning token limit.
If `thinking_token_budget` is not specified, no explicit reasoning limit is applied beyond normal generation constraints such as `max_tokens`.
@@ -257,20 +257,20 @@ If `thinking_token_budget` is not specified, no explicit reasoning limit is appl
`--reasoning-config` accepts a JSON object corresponding to
[ReasoningConfig][vllm.config.ReasoningConfig] with the following fields:
| Field | Type | Description |
|-------------------|----------------|--------------------------------------------------|
| `think_start_str` | `str \| null` | String that marks the start of reasoning content |
| `think_end_str` | `str \| null` | String that marks the end of reasoning content |
| Field | Type | Description |
|-----------------------|----------------|--------------------------------------------------|
| `reasoning_start_str` | `str \| null` | String that marks the start of reasoning content |
| `reasoning_end_str` | `str \| null` | String that marks the end of reasoning content |
!!! note
`think_end_str` can include a transition phrase before the think end token. For example, setting `think_end_str` to `"I have to give the solution based on the thinking directly now.</think>"` instructs the model to emit that phrase when the budget is exhausted, making the reasoning termination more natural.
`reasoning_end_str` can include a transition phrase before the reasoning end token. For example, setting `reasoning_end_str` to `"I have to give the solution based on the reasoning directly now.</think>"` instructs the model to emit that phrase when the budget is exhausted, making the reasoning termination more natural.
### Online Serving
```bash
vllm serve Qwen/Qwen3-0.6B \
--reasoning-parser qwen3 \
--reasoning-config '{"think_start_str": "<think>", "think_end_str": "I have to give the solution based on the thinking directly now.</think>"}'
--reasoning-config '{"reasoning_start_str": "<think>", "reasoning_end_str": "I have to give the solution based on the reasoning directly now.</think>"}'
```
Then make a request with `thinking_token_budget` to limit the reasoning tokens:
@@ -298,8 +298,8 @@ from vllm.config import ReasoningConfig
llm = LLM(
model="Qwen/Qwen3-0.6B",
reasoning_config=ReasoningConfig(
think_start_str="<think>",
think_end_str="I have to give the solution based on the thinking directly now.</think>",
reasoning_start_str="<think>",
reasoning_end_str="I have to give the solution based on the thinking directly now.</think>",
),
)

2
tests/v1/entrypoints/openai/test_thinking_token_budget.py
View File

@@ -20,7 +20,7 @@ def server():
"--reasoning-parser",
"qwen3",
"--reasoning-config",
'{"think_start_str": "<think>", "think_end_str": "</think>"}',
'{"reasoning_start_str": "<think>", "reasoning_end_str": "</think>"}',
"--max-model-len",
"2048",
"--enforce-eager",

8
tests/v1/logits_processors/test_correctness.py
View File

@@ -103,8 +103,8 @@ class LogitsProcsRequestParams:
class MockReasoningConfig:
"""Mock reasoning config for testing ThinkingTokenBudgetLogitsProcessor."""
think_start_token_ids = [THINK_START_TOKEN_ID]
think_end_token_ids = [THINK_END_TOKEN_ID]
reasoning_start_token_ids = [THINK_START_TOKEN_ID]
reasoning_end_token_ids = [THINK_END_TOKEN_ID]
def _generate_fake_sampling_metadata(
@@ -491,7 +491,7 @@ def _thinking_budget_validate(
# Find if thinking has started in output tokens
thinking_started = False
start_tokens = tb_processor.think_start_token_ids
start_tokens = tb_processor.reasoning_start_token_ids
if len(start_tokens) > 0:
for i in range(len(output_tokens) - len(start_tokens) + 1):
@@ -518,7 +518,7 @@ def _thinking_budget_validate(
)
# Validate that only end tokens are allowed
end_tokens = tb_processor.think_end_token_ids
end_tokens = tb_processor.reasoning_end_token_ids
if len(end_tokens) > 0:
expected_end_token_id = end_tokens[
min(state["end_count"], len(end_tokens) - 1)

46
vllm/config/reasoning.py
View File

@@ -12,7 +12,7 @@ from vllm.tokenizers import cached_tokenizer_from_config
class ReasoningConfig:
"""Configuration for reasoning models.
Set `think_start_str` and `think_end_str` to the strings that delimit
Set `reasoning_start_str` and `reasoning_end_str` to the strings that delimit
the reasoning block (e.g. `"<think>"` and `"</think>"`). The
corresponding token IDs are derived automatically via
`initialize_token_ids` and are not intended to be set directly.
@@ -20,53 +20,55 @@ class ReasoningConfig:
# NOTE: These parameters are temporary, the intent is to derive them
# automatically from the reasoning parser in a future version.
think_start_str: str = "<think>"
reasoning_start_str: str = "<think>"
"""String that indicates the start of reasoning."""
think_end_str: str = "</think>"
reasoning_end_str: str = "</think>"
"""String that indicates the end of reasoning content."""
_think_start_token_ids: list[int] | None = field(
_reasoning_start_token_ids: list[int] | None = field(
default=None, init=False, repr=False
)
"""Private backing field for `think_start_token_ids`. Set by
"""Private backing field for `reasoning_start_token_ids`. Set by
`initialize_token_ids`. Not intended to be configured directly."""
_think_end_token_ids: list[int] | None = field(default=None, init=False, repr=False)
"""Private backing field for `think_end_token_ids`. Set by
_reasoning_end_token_ids: list[int] | None = field(
default=None, init=False, repr=False
)
"""Private backing field for `reasoning_end_token_ids`. Set by
`initialize_token_ids`. Not intended to be configured directly."""
@property
def think_start_token_ids(self) -> list[int] | None:
"""Token IDs derived from `think_start_str`. Set automatically by
def reasoning_start_token_ids(self) -> list[int] | None:
"""Token IDs derived from `reasoning_start_str`. Set automatically by
`initialize_token_ids`. Not intended to be configured directly."""
return self._think_start_token_ids
return self._reasoning_start_token_ids
@property
def think_end_token_ids(self) -> list[int] | None:
"""Token IDs derived from `think_end_str`. Set automatically by
def reasoning_end_token_ids(self) -> list[int] | None:
"""Token IDs derived from `reasoning_end_str`. Set automatically by
`initialize_token_ids`. Not intended to be configured directly."""
return self._think_end_token_ids
return self._reasoning_end_token_ids
def initialize_token_ids(self, model_config: ModelConfig) -> None:
"""Initialize reasoning token IDs from strings using the tokenizer."""
if (
self._think_start_token_ids is not None
and self._think_end_token_ids is not None
self._reasoning_start_token_ids is not None
and self._reasoning_end_token_ids is not None
):
return
tokenizer = cached_tokenizer_from_config(model_config=model_config)
self._think_start_token_ids = tokenizer.encode(
self.think_start_str, add_special_tokens=False
self._reasoning_start_token_ids = tokenizer.encode(
self.reasoning_start_str, add_special_tokens=False
)
self._think_end_token_ids = tokenizer.encode(
self.think_end_str, add_special_tokens=False
self._reasoning_end_token_ids = tokenizer.encode(
self.reasoning_end_str, add_special_tokens=False
)
if not self._think_start_token_ids or not self._think_end_token_ids:
if not self._reasoning_start_token_ids or not self._reasoning_end_token_ids:
raise ValueError(
f"ReasoningConfig: failed to tokenize reasoning strings: "
f"think_start_str='{self.think_start_str}', "
f"think_end_str='{self.think_end_str}'. "
f"reasoning_start_str='{self.reasoning_start_str}', "
f"reasoning_end_str='{self.reasoning_end_str}'. "
"Ensure the strings are valid tokens in the model's vocabulary."
)

28
vllm/v1/sample/logits_processor/builtin.py
View File

@@ -303,10 +303,12 @@ class ThinkingTokenBudgetLogitsProcessor(LogitsProcessor):
# Check if thinking is enabled
self.is_enabled = reasoning_config is not None
self.think_start_token_ids = getattr(
reasoning_config, "think_start_token_ids", []
self.reasoning_start_token_ids = getattr(
reasoning_config, "reasoning_start_token_ids", []
)
self.reasoning_end_token_ids = getattr(
reasoning_config, "reasoning_end_token_ids", []
)
self.think_end_token_ids = getattr(reasoning_config, "think_end_token_ids", [])
self.pin_memory = is_pin_memory
self.device = device
@@ -357,15 +359,15 @@ class ThinkingTokenBudgetLogitsProcessor(LogitsProcessor):
think_count = 0
else:
last_start = self._find_last_sequence_index(
prompt_tok_ids, self.think_start_token_ids
prompt_tok_ids, self.reasoning_start_token_ids
)
last_end = self._find_last_sequence_index(
prompt_tok_ids, self.think_end_token_ids
prompt_tok_ids, self.reasoning_end_token_ids
)
in_think = last_start > last_end
if in_think:
think_count = len(prompt_tok_ids) - (
last_start + len(self.think_start_token_ids)
last_start + len(self.reasoning_start_token_ids)
)
else:
think_count = 0
@@ -405,8 +407,8 @@ class ThinkingTokenBudgetLogitsProcessor(LogitsProcessor):
state["prev_output_length"] = current_length
# Check if new tokens contain think start or end sequences
start_len = len(self.think_start_token_ids)
end_len = len(self.think_end_token_ids)
start_len = len(self.reasoning_start_token_ids)
end_len = len(self.reasoning_end_token_ids)
# Look for think sequences in recent tokens (including boundary)
# Check overlapping regions where sequences might span boundaries
@@ -415,10 +417,10 @@ class ThinkingTokenBudgetLogitsProcessor(LogitsProcessor):
# Find any think start/end sequences in recent tokens
recent_start_pos = self._find_last_sequence_index(
recent_tokens, self.think_start_token_ids
recent_tokens, self.reasoning_start_token_ids
)
recent_end_pos = self._find_last_sequence_index(
recent_tokens, self.think_end_token_ids
recent_tokens, self.reasoning_end_token_ids
)
# Update state based on recent sequences
@@ -469,7 +471,7 @@ class ThinkingTokenBudgetLogitsProcessor(LogitsProcessor):
else:
# In end mode
state["end_count"] += 1
if state["end_count"] >= len(self.think_end_token_ids):
if state["end_count"] >= len(self.reasoning_end_token_ids):
state.update(
{
"in_end": False,
@@ -530,7 +532,9 @@ class ThinkingTokenBudgetLogitsProcessor(LogitsProcessor):
state = self._state.get(i)
if state and state["in_end"]:
self.mask[i] = True
self.force_token_ids[i] = self.think_end_token_ids[state["end_count"]]
self.force_token_ids[i] = self.reasoning_end_token_ids[
state["end_count"]
]
# Check in CPU first not to sync with GPU
has_active_thinking = any(