File size: 9,743 Bytes
0175530 d1ee588 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 | import json
from collections.abc import Sequence
from typing import Any, Optional
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.engine.protocol import (
DeltaFunctionCall,
DeltaMessage,
DeltaToolCall,
ExtractedToolCallInformation,
FunctionCall,
ToolCall,
)
from vllm.tokenizers import TokenizerLike
from vllm.tool_parsers.abstract_tool_parser import ToolParser, ToolParserManager
@ToolParserManager.register_module(["openpipe_llama_dual"])
class OpenPipeLlamaDualParser(ToolParser):
"""Parse official JSON, llama31 tool markers, and pipeline3 function tags."""
LEGACY_START = "<|start_tool_call|>"
LEGACY_END = "<|end_tool_call|>"
FUNCTION_CALL_TAG = "<function>"
FUNCTION_ARGS_TAG = "<arguments>"
VARIANT_LLAMA31 = "llama31instruct"
VARIANT_PIPELINE3 = "pipeline3"
VARIANT_OFFICIAL = "official"
def __init__(self, tokenizer: TokenizerLike, tools):
super().__init__(tokenizer, tools)
self.tokenizer = tokenizer
self.tools = tools
def _get_template_variant(self, request: ChatCompletionRequest) -> Optional[str]:
kwargs = getattr(request, "chat_template_kwargs", None)
if kwargs is None:
return None
if isinstance(kwargs, dict):
value = kwargs.get("template_variant")
return value if isinstance(value, str) else None
value = getattr(kwargs, "template_variant", None)
return value if isinstance(value, str) else None
def _normalize_tool_call(self, payload: dict[str, Any]) -> Optional[dict[str, Any]]:
if "name" in payload and "parameters" in payload:
return {
"name": payload["name"],
"arguments": payload["parameters"],
}
if "function" in payload and isinstance(payload["function"], dict):
function = payload["function"]
if "name" in function and "arguments" in function:
return {
"name": function["name"],
"arguments": function["arguments"],
}
return None
def _extract_legacy_tool_calls(self, text: str) -> list[dict[str, Any]]:
tool_calls = []
current_index = 0
while True:
start_index = text.find(self.LEGACY_START, current_index)
if start_index == -1:
break
end_index = text.find(self.LEGACY_END, start_index)
if end_index == -1:
break
tool_call_json = text[start_index + len(self.LEGACY_START) : end_index].strip()
payload = json.loads(tool_call_json)
normalized = self._normalize_tool_call(payload)
if normalized:
tool_calls.append(normalized)
current_index = end_index + len(self.LEGACY_END)
return tool_calls
def _extract_function_tag_tool_calls(self, text: str) -> list[dict[str, Any]]:
tool_calls = []
current_index = 0
while True:
function_start = text.find(self.FUNCTION_CALL_TAG, current_index)
if function_start == -1:
break
name_start = function_start + len(self.FUNCTION_CALL_TAG)
args_tag_index = text.find(self.FUNCTION_ARGS_TAG, name_start)
if args_tag_index == -1:
break
function_name = text[name_start:args_tag_index].strip()
if not function_name:
break
arguments_start = args_tag_index + len(self.FUNCTION_ARGS_TAG)
next_function_index = text.find(self.FUNCTION_CALL_TAG, arguments_start)
if next_function_index == -1:
arguments_raw = text[arguments_start:].strip()
current_index = len(text)
else:
arguments_raw = text[arguments_start:next_function_index].strip()
current_index = next_function_index
if not arguments_raw:
arguments: Any = ""
else:
try:
arguments = json.loads(arguments_raw)
except Exception:
arguments = arguments_raw
tool_calls.append(
{
"name": function_name,
"arguments": arguments,
}
)
return tool_calls
def _extract_official_tool_call(self, text: str) -> Optional[dict[str, Any]]:
stripped = text.strip()
if not stripped.startswith("{") or not stripped.endswith("}"):
return None
payload = json.loads(stripped)
return self._normalize_tool_call(payload)
def _build_delta_tool_call(self, tool_call: dict[str, Any], index: int = 0) -> DeltaMessage:
arguments = tool_call["arguments"]
return DeltaMessage(
tool_calls=[
DeltaToolCall(
index=index,
id=f"call_{tool_call['name']}",
type="function",
function=DeltaFunctionCall(
name=tool_call["name"],
arguments=json.dumps(arguments, ensure_ascii=False)
if isinstance(arguments, (dict, list))
else arguments,
),
)
]
)
def _build_tool_calls_response(
self,
tool_calls: list[dict[str, Any]],
) -> ExtractedToolCallInformation:
return ExtractedToolCallInformation(
tools_called=True,
tool_calls=[
ToolCall(
id=f"call_{index + 1}",
type="function",
function=FunctionCall(
name=tool_call["name"],
arguments=json.dumps(
tool_call["arguments"], ensure_ascii=False
)
if isinstance(tool_call["arguments"], (dict, list))
else tool_call["arguments"],
),
)
for index, tool_call in enumerate(tool_calls)
],
content=None,
)
def _looks_like_partial_official_json(self, text: str) -> bool:
stripped = text.strip()
if not stripped.startswith("{"):
return False
if stripped.endswith("}"):
return False
return (
'"name"' in stripped
or '"parameters"' in stripped
or '"function"' in stripped
)
def extract_tool_calls_streaming(
self,
previous_text: str,
current_text: str,
delta_text: str,
previous_token_ids: Sequence[int],
current_token_ids: Sequence[int],
delta_token_ids: Sequence[int],
request: ChatCompletionRequest,
) -> DeltaMessage | None:
variant = self._get_template_variant(request)
try:
if (
variant == self.VARIANT_LLAMA31
or self.LEGACY_START in current_text
):
if self.LEGACY_START in current_text and self.LEGACY_END in current_text:
tool_calls = self._extract_legacy_tool_calls(current_text)
if tool_calls:
return self._build_delta_tool_call(
tool_calls[-1], index=len(tool_calls) - 1
)
if self.LEGACY_START in current_text:
return None
return DeltaMessage(content=delta_text)
if variant == self.VARIANT_PIPELINE3 or self.FUNCTION_CALL_TAG in current_text:
tool_calls = self._extract_function_tag_tool_calls(current_text)
if tool_calls:
return self._build_delta_tool_call(
tool_calls[-1], index=len(tool_calls) - 1
)
return None
official_tool_call = self._extract_official_tool_call(current_text)
if official_tool_call:
return self._build_delta_tool_call(official_tool_call)
if variant == self.VARIANT_OFFICIAL and self._looks_like_partial_official_json(
current_text
):
return None
except Exception:
return DeltaMessage(content=delta_text)
return DeltaMessage(content=delta_text)
def extract_tool_calls(
self,
model_output: str,
request: ChatCompletionRequest,
) -> ExtractedToolCallInformation:
variant = self._get_template_variant(request)
try:
if (
variant == self.VARIANT_LLAMA31
or self.LEGACY_START in model_output
):
tool_calls = self._extract_legacy_tool_calls(model_output)
if tool_calls:
return self._build_tool_calls_response(tool_calls)
if variant == self.VARIANT_PIPELINE3 or self.FUNCTION_CALL_TAG in model_output:
tool_calls = self._extract_function_tag_tool_calls(model_output)
if tool_calls:
return self._build_tool_calls_response(tool_calls)
official_tool_call = self._extract_official_tool_call(model_output)
if official_tool_call:
return self._build_tool_calls_response([official_tool_call])
except Exception:
pass
return ExtractedToolCallInformation(
tools_called=False,
tool_calls=[],
content=model_output,
) |