llama4_pythonic_tool_parser.py 12.5 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
import ast
import json
from collections.abc import Sequence
6
from typing import Any
7

8
import regex as re
9
10
from transformers import PreTrainedTokenizerBase

11
import vllm.envs as envs
12
13
14
15
16
17
18
19
20
from vllm.entrypoints.openai.protocol import (
    ChatCompletionRequest,
    DeltaFunctionCall,
    DeltaMessage,
    DeltaToolCall,
    ExtractedToolCallInformation,
    FunctionCall,
    ToolCall,
)
21
22
from vllm.logger import init_logger
from vllm.tool_parsers.abstract_tool_parser import (
23
24
    ToolParser,
)
25
26
27
28
29
30
31
32
33
34
35
36
37

logger = init_logger(__name__)


class _UnexpectedAstError(Exception):
    pass


class Llama4PythonicToolParser(ToolParser):
    """
    Toolcall parser for Llama4 that produce tool calls in a pythonic style
    Use --enable-auto-tool-choice --tool-call-parser llama4_pythonic
    """
38

39
40
41
42
43
44
45
46
47
    # TODO(mdepinet): Possible future improvements:
    #   1. Support text + tools separated by either <|python_tag|> or \n\n
    #   2. Support tools outside of a list (or separated by a semicolon).
    #      This depends on item 1 for consistent streaming.
    # Neither of these are necessary for e.g. ToolACE, but both would help make
    # Llama3.2 models more reliable.

    TOOL_CALL_REGEX = re.compile(
        r"\[([a-zA-Z]+\w*\(([a-zA-Z]+\w*=.*,\s*)*([a-zA-Z]+\w*=.*\s)?\),\s*)*([a-zA-Z]+\w*\(([a-zA-Z]+\w*=.*,\s*)*([a-zA-Z]+\w*=.*\s*)?\)\s*)+\]",
48
49
        re.DOTALL,
    )
50
51
52
53
54
55
56
57
58
59
60
61
62
63

    def __init__(self, tokenizer: PreTrainedTokenizerBase):
        super().__init__(tokenizer)

    # Rename for readability. This is NOT a tool id.
    @property
    def current_tool_index(self) -> int:
        return self.current_tool_id

    @current_tool_index.setter
    def current_tool_index(self, value: int) -> None:
        self.current_tool_id = value

    def extract_tool_calls(
64
65
        self, model_output: str, request: ChatCompletionRequest
    ) -> ExtractedToolCallInformation:
66
67
68
69
70
71
72
        """
        Extract the tool calls from a complete model response.
        """

        # remove <|python_start|> and <|python_end|>
        # as Llama 4 model sometime will output those tokens
        if model_output.startswith("<|python_start|>"):
73
            model_output = model_output[len("<|python_start|>") :]
74
            model_output = model_output.replace("<|python_end|>", "")
75
76
77

        is_tool_call_pattern = False
        try:
78
79
80
81
82
83
            is_tool_call_pattern = (
                self.TOOL_CALL_REGEX.match(
                    model_output, timeout=envs.VLLM_TOOL_PARSE_REGEX_TIMEOUT_SECONDS
                )
                is not None
            )
84
        except TimeoutError:
85
86
87
88
            logger.warning("Regex timeout occurred when matching tool call pattern.")
            logger.debug(
                "Regex timeout occurred when matching user input: %s", model_output
            )
89
90

        if not is_tool_call_pattern:
91
92
93
            return ExtractedToolCallInformation(
                tools_called=False, tool_calls=[], content=model_output
            )
94
95
96
97
98

        try:
            module = ast.parse(model_output)
            parsed = getattr(module.body[0], "value", None)
            if isinstance(parsed, ast.List) and all(
99
100
                isinstance(e, ast.Call) for e in parsed.elts
            ):
101
102
103
104
105
106
                return ExtractedToolCallInformation(
                    tools_called=True,
                    tool_calls=[
                        _handle_single_tool(e)  # type: ignore
                        for e in parsed.elts
                    ],
107
108
                    content=None,
                )
109
110
            else:
                raise _UnexpectedAstError(
111
112
                    "Tool output must be a list of function calls"
                )
113
114
115
        except Exception:
            logger.exception("Error in extracting tool call from response.")
            # Treat as regular text
116
117
118
            return ExtractedToolCallInformation(
                tools_called=False, tool_calls=[], content=model_output
            )
119
120
121
122
123
124
125
126
127
128

    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,
129
    ) -> DeltaMessage | None:
130
        if not current_text.startswith("[") and not current_text.startswith(
131
132
            "<|python_start|>"
        ):
133
134
135
136
137
            return DeltaMessage(content=delta_text)

        try:
            # remove <|python_start|> and <|python_end|>
            if current_text.startswith("<|python_start|>"):
138
                current_text = current_text[len("<|python_start|>") :]
139
            if current_text.endswith("<|python_end|>"):
140
                current_text = current_text[: current_text.rfind("<|python_end|>")]
141
142
143
144
145
146
147
148
            valid_and_added_text = _make_valid_python(current_text)
            if valid_and_added_text is None:
                return None
            valid_text, added_text = valid_and_added_text

            module = ast.parse(valid_text)
            parsed = getattr(module.body[0], "value", None)
            if not isinstance(parsed, ast.List) or not all(
149
150
                isinstance(e, ast.Call) for e in parsed.elts
            ):
151
                raise _UnexpectedAstError(
152
153
                    "Tool output must be a list of function calls"
                )
154
155
156
157
158
159
160
161
162
163
164
165
166
167
            tool_calls = [
                _handle_single_tool(e)  # type: ignore
                for e in parsed.elts
            ]

            tool_deltas = []
            for index, new_call in enumerate(tool_calls):
                if index < self.current_tool_index:
                    continue

                self.current_tool_index = index
                if len(self.streamed_args_for_tool) == index:
                    self.streamed_args_for_tool.append("")

168
169
170
                new_call_complete = (
                    index < len(tool_calls) - 1 or ")]" not in added_text
                )
171
172
173
                if new_call_complete:
                    self.current_tool_index += 1

174
                withheld_suffix = added_text[:-2] if not new_call_complete else ""
175
176
177
178
179
180
                if not new_call_complete and added_text[-2] == ")":
                    # Function call is incomplete. Withhold the closing bracket.
                    withheld_suffix = withheld_suffix + "}"
                # Strings get single quotes in the model-produced string.
                # JSON requires double quotes.
                withheld_suffix = withheld_suffix.replace("'", '"')
181
182
183
                delta = _compute_tool_delta(
                    self.streamed_args_for_tool[index], new_call, index, withheld_suffix
                )
184
185
186

                if delta is not None:
                    tool_deltas.append(delta)
187
188
189
190
191
192
193
194
195
196
197
                    if (
                        delta.function is not None
                        and delta.function.arguments is not None
                    ):
                        self.streamed_args_for_tool[index] += delta.function.arguments

            # HACK: serving_chat.py inspects the internal state of tool parsers
            # when determining its final streaming delta, automatically
            # adding autocompleted JSON.
            # These two lines avoid that nonsense while ensuring finish_reason
            # is set to tool_calls when at least one tool is called.
198
199
200
201
202
203
204
205
            if tool_deltas and not self.prev_tool_call_arr:
                self.prev_tool_call_arr = [{"arguments": {}}]

            if tool_deltas:
                return DeltaMessage(tool_calls=tool_deltas)
            elif not added_text and self.current_tool_id > 0:
                # Return an empty DeltaMessage once the tool calls are all done
                # so that finish_reason gets set.
206
                return DeltaMessage(content="")
207
208
209
210
211
            else:
                return None
        except Exception:
            logger.exception("Error trying to handle streaming tool call.")
            logger.debug(
212
213
                "Skipping chunk as a result of tool streaming extraction error"
            )
214
215
216
217
218
219
220
221
            return None


def _get_parameter_value(val: ast.expr) -> Any:
    if isinstance(val, ast.Constant):
        return val.value
    elif isinstance(val, ast.Dict):
        if not all(isinstance(k, ast.Constant) for k in val.keys):
222
            raise _UnexpectedAstError("Dict tool call arguments must have literal keys")
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
        return {
            k.value: _get_parameter_value(v)  # type: ignore
            for k, v in zip(val.keys, val.values)
        }
    elif isinstance(val, ast.List):
        return [_get_parameter_value(v) for v in val.elts]
    else:
        raise _UnexpectedAstError("Tool call arguments must be literals")


def _handle_single_tool(call: ast.Call) -> ToolCall:
    if not isinstance(call.func, ast.Name):
        raise _UnexpectedAstError("Invalid tool call name")
    function_name = call.func.id
    arguments = {}
    for keyword in call.keywords:
        arguments[keyword.arg] = _get_parameter_value(keyword.value)
240
241
242
243
    return ToolCall(
        type="function",
        function=FunctionCall(name=function_name, arguments=json.dumps(arguments)),
    )
244
245


246
def _make_valid_python(text: str) -> tuple[str, str] | None:
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
    bracket_stack = []
    for index, char in enumerate(text):
        if char in {"[", "(", "{"}:
            bracket_stack.append(char)
        elif char == "]":
            if not bracket_stack or bracket_stack.pop() != "[":
                raise _UnexpectedAstError("Mismatched square brackets")
        elif char == ")":
            if not bracket_stack or bracket_stack.pop() != "(":
                raise _UnexpectedAstError("Mismatched parentheses")
        elif char == "}":
            if not bracket_stack or bracket_stack.pop() != "{":
                raise _UnexpectedAstError("Mismatched curly braces")
        elif char in {"'", '"'}:
            if bracket_stack and bracket_stack[-1] == char:
                if index > 0 and text[index - 1] == "\\":
                    # Treat an escaped quote as a regular character
                    pass
                else:
                    bracket_stack.pop()
            elif bracket_stack and bracket_stack[-1] in {"'", '"'}:
                # Double quote within a single quote string or vice versa.
                pass
            else:
                bracket_stack.append(char)

    text = text.rstrip()
    if text.endswith("=") or text.endswith(":"):
        # Since we have no type information for this property/parameter value,
        # we can't fill in a valid value.
        return None
    if bracket_stack and bracket_stack[-1] == "{":
279
        trailing_dict_text = text[: text.rfind("{")]
280
281
282
283
284
        num_keys = trailing_dict_text.count(":")
        num_values = trailing_dict_text.count(",")
        if num_keys <= num_values:
            return None  # Incomplete property name within parameter value
    if bracket_stack and bracket_stack[-1] == "(":
285
        trailing_params_text = text[: text.rfind("(")]
286
287
288
289
290
291
        num_full_param_names = trailing_params_text.count("=")
        num_full_param_values = trailing_params_text.count(",")
        if num_full_param_names <= num_full_param_values:
            return None  # Incomplete parameter name
    if text.endswith(","):
        text = text[:-1]
292
293
294
295
296
297
    if (
        bracket_stack
        and bracket_stack[-1] == "["
        and not text.endswith("[")
        and not text.endswith(")")
    ):
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
        return None  # Incomplete function name

    added_text = ""
    for char in reversed(bracket_stack):
        if char == "[":
            added_text += "]"
        elif char == "(":
            added_text += ")"
        elif char == "{":
            added_text += "}"
        elif char == "'":
            added_text += "'"
        elif char == '"':
            added_text += '"'

    return text + added_text, added_text


316
317
def _compute_tool_delta(
    previously_sent_args: str, new_call: ToolCall, index: int, withheld_suffix: str
318
) -> DeltaToolCall | None:
319
320
321
    new_call_args = new_call.function.arguments
    if withheld_suffix:
        assert new_call_args.endswith(withheld_suffix)
322
        new_call_args = new_call_args[: -len(withheld_suffix)]
323
    if not previously_sent_args:
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
        return DeltaToolCall(
            id=new_call.id,
            type="function",
            index=index,
            function=DeltaFunctionCall(
                name=new_call.function.name,
                arguments=new_call_args,
            ),
        )

    arg_diff = new_call_args[len(previously_sent_args) :]
    return (
        DeltaToolCall(
            id=None, index=index, function=DeltaFunctionCall(arguments=arg_diff)
        )
        if arg_diff
        else None
    )