[Chore] Cleanup guided namespace, move to structured outputs config (#22772)

Signed-off-by: Aaron Pham <contact@aarnphm.xyz>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Aaron Pham
2025-09-18 09:20:27 +00:00
committed by GitHub
co-authored by Harry Mellor
parent 05b044e698
commit 29283e8976
51 changed files with 579 additions and 806 deletions
@@ -1,11 +1,10 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
This file demonstrates the example usage of guided decoding
to generate structured outputs using vLLM. It shows how to apply
different guided decoding techniques such as Choice, Regex, JSON schema,
and Grammar to produce structured and formatted results
based on specific prompts.
This file demonstrates the example usage of structured outputs
in vLLM. It shows how to apply different constraints such as choice,
regex, json schema, and grammar to produce structured and formatted
results based on specific prompts.
"""
from enum import Enum
@@ -13,19 +12,23 @@ from enum import Enum
from pydantic import BaseModel
from vllm import LLM, SamplingParams
from vllm.sampling_params import GuidedDecodingParams
from vllm.sampling_params import StructuredOutputsParams
MAX_TOKENS = 50
# Guided decoding by Choice (list of possible options)
guided_decoding_params_choice = GuidedDecodingParams(choice=["Positive", "Negative"])
sampling_params_choice = SamplingParams(guided_decoding=guided_decoding_params_choice)
# Structured outputs by Choice (list of possible options)
structured_outputs_params_choice = StructuredOutputsParams(
choice=["Positive", "Negative"]
)
sampling_params_choice = SamplingParams(
structured_outputs=structured_outputs_params_choice
)
prompt_choice = "Classify this sentiment: vLLM is wonderful!"
# Guided decoding by Regex
guided_decoding_params_regex = GuidedDecodingParams(regex=r"\w+@\w+\.com\n")
# Structured outputs by Regex
structured_outputs_params_regex = StructuredOutputsParams(regex=r"\w+@\w+\.com\n")
sampling_params_regex = SamplingParams(
guided_decoding=guided_decoding_params_regex,
structured_outputs=structured_outputs_params_regex,
stop=["\n"],
max_tokens=MAX_TOKENS,
)
@@ -36,7 +39,7 @@ prompt_regex = (
)
# Guided decoding by JSON using Pydantic schema
# Structured outputs by JSON using Pydantic schema
class CarType(str, Enum):
sedan = "sedan"
suv = "SUV"
@@ -51,17 +54,16 @@ class CarDescription(BaseModel):
json_schema = CarDescription.model_json_schema()
guided_decoding_params_json = GuidedDecodingParams(json=json_schema)
structured_outputs_params_json = StructuredOutputsParams(json=json_schema)
sampling_params_json = SamplingParams(
guided_decoding=guided_decoding_params_json,
max_tokens=MAX_TOKENS,
structured_outputs=structured_outputs_params_json, max_tokens=MAX_TOKENS
)
prompt_json = (
"Generate a JSON with the brand, model and car_type of"
"Generate a JSON with the brand, model and car_type of "
"the most iconic car from the 90's"
)
# Guided decoding by Grammar
# Structured outputs by Grammar
simplified_sql_grammar = """
root ::= select_statement
select_statement ::= "SELECT " column " from " table " where " condition
@@ -70,13 +72,15 @@ table ::= "table_1 " | "table_2 "
condition ::= column "= " number
number ::= "1 " | "2 "
"""
guided_decoding_params_grammar = GuidedDecodingParams(grammar=simplified_sql_grammar)
structured_outputs_params_grammar = StructuredOutputsParams(
grammar=simplified_sql_grammar
)
sampling_params_grammar = SamplingParams(
guided_decoding=guided_decoding_params_grammar,
structured_outputs=structured_outputs_params_grammar,
max_tokens=MAX_TOKENS,
)
prompt_grammar = (
"Generate an SQL query to show the 'username' and 'email'from the 'users' table."
"Generate an SQL query to show the 'username' and 'email' from the 'users' table."
)
@@ -93,16 +97,16 @@ def main():
llm = LLM(model="Qwen/Qwen2.5-3B-Instruct", max_model_len=100)
choice_output = generate_output(prompt_choice, sampling_params_choice, llm)
format_output("Guided decoding by Choice", choice_output)
format_output("Structured outputs by Choice", choice_output)
regex_output = generate_output(prompt_regex, sampling_params_regex, llm)
format_output("Guided decoding by Regex", regex_output)
format_output("Structured outputs by Regex", regex_output)
json_output = generate_output(prompt_json, sampling_params_json, llm)
format_output("Guided decoding by JSON", json_output)
format_output("Structured outputs by JSON", json_output)
grammar_output = generate_output(prompt_grammar, sampling_params_grammar, llm)
format_output("Guided decoding by Grammar", grammar_output)
format_output("Structured outputs by Grammar", grammar_output)
if __name__ == "__main__":
@@ -6,7 +6,7 @@ without any specific flags:
```bash
VLLM_USE_V1=0 vllm serve unsloth/Llama-3.2-1B-Instruct \
--guided-decoding-backend outlines
--structured-outputs-config.backend outlines
```
This example demonstrates how to generate chat completions
@@ -86,7 +86,7 @@ PARAMS: dict[ConstraintsFormat, dict[str, Any]] = {
"content": "Classify this sentiment: vLLM is wonderful!",
}
],
"extra_body": {"guided_choice": ["positive", "negative"]},
"extra_body": {"structured_outputs": {"choice": ["positive", "negative"]}},
},
"regex": {
"messages": [
@@ -96,7 +96,7 @@ PARAMS: dict[ConstraintsFormat, dict[str, Any]] = {
}
],
"extra_body": {
"guided_regex": r"[a-z0-9.]{1,20}@\w{6,10}\.com\n",
"structured_outputs": {"regex": r"[a-z0-9.]{1,20}@\w{6,10}\.com\n"},
},
},
"json": {
@@ -122,7 +122,8 @@ PARAMS: dict[ConstraintsFormat, dict[str, Any]] = {
}
],
"extra_body": {
"guided_grammar": """
"structured_outputs": {
"grammar": """
root ::= select_statement
select_statement ::= "SELECT " column " from " table " where " condition
@@ -135,6 +136,7 @@ condition ::= column "= " number
number ::= "1 " | "2 "
""",
}
},
},
"structural_tag": {