Skip to content

vllm.tool_parsers.poolside_v1_tool_parser

GLM-4 Tool Call Parser with incremental string streaming support.

This parser fixes the streaming issue reported in Issue #32829 where long string parameters (e.g., file content with 4000+ characters of code) are buffered until complete, causing multi-second delays before the user sees any content.

The fix streams string values incrementally as they arrive, providing a true streaming experience for long content.

PoolsideV1ToolParser

Bases: ToolParser

Tool parser for GLM-4 models with incremental string streaming.

This parser emits tool-call deltas incrementally as arguments arrive. For string-type parameters, content is streamed character-by-character rather than waiting for the complete tag.

Source code in vllm/tool_parsers/poolside_v1_tool_parser.py
 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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
class PoolsideV1ToolParser(ToolParser):
    """Tool parser for GLM-4 models with incremental string streaming.

    This parser emits tool-call deltas incrementally as arguments arrive.
    For string-type parameters, content is streamed character-by-character
    rather than waiting for the complete </arg_value> tag.
    """

    def __init__(self, tokenizer: TokenizerLike, tools: list[Tool] | None = None):
        super().__init__(tokenizer, tools)
        # Stateful streaming fields
        self.current_tool_name_sent: bool = False
        self.prev_tool_call_arr: list[dict[str, Any]] = []
        self.current_tool_id: int = -1
        self.streamed_args_for_tool: list[str] = []

        self.tool_call_start_token: str = "<tool_call>"
        self.tool_call_end_token: str = "</tool_call>"
        self.arg_key_start: str = "<arg_key>"
        self.arg_key_end: str = "</arg_key>"
        self.arg_val_start: str = "<arg_value>"
        self.arg_val_end: str = "</arg_value>"

        self.tool_calls_start_token = self.tool_call_start_token

        self.func_call_regex = re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL)
        self.func_detail_regex = re.compile(
            r"<tool_call>([^\n]*)\n(.*)</tool_call>", re.DOTALL
        )
        self.func_arg_regex = re.compile(
            r"<arg_key>(.*?)</arg_key>\s*<arg_value>(.*?)</arg_value>", re.DOTALL
        )

        if not self.model_tokenizer:
            raise ValueError(
                "The model tokenizer must be passed to the ToolParser "
                "constructor during construction."
            )

        self.tool_call_start_token_id = self.vocab.get(self.tool_call_start_token)
        self.tool_call_end_token_id = self.vocab.get(self.tool_call_end_token)
        self._buffer: str = ""

        # Streaming state for incremental tool-call streaming
        self._in_tool_call: bool = False
        self._current_tool_name: str | None = None
        self._pending_key: str | None = None
        self._streaming_string_value: bool = False
        self._tool_call_ids: list[str] = []
        self._args_started: list[bool] = []
        self._args_closed: list[bool] = []
        self._seen_keys: list[set[str]] = []

    @staticmethod
    def _deserialize(value: str) -> Any:
        try:
            return json.loads(value)
        except json.JSONDecodeError:
            pass

        try:
            return ast.literal_eval(value)
        except (ValueError, SyntaxError):
            pass

        return value

    @staticmethod
    def _json_escape_string_content(s: str) -> str:
        """JSON-escape string content for incremental streaming.

        This escapes the content that goes INSIDE a JSON string (between quotes),
        not including the surrounding quotes themselves.
        """
        if not s:
            return ""
        return json.dumps(s, ensure_ascii=False)[1:-1]

    @staticmethod
    def _is_string_type(
        tool_name: str,
        arg_name: str,
        tools: list[Tool] | None,
    ) -> bool:
        if tools is None:
            return False
        for tool in tools:
            if tool.function.name != tool_name:
                continue
            if tool.function.parameters is None:
                return False
            arg_type = (
                tool.function.parameters.get("properties", {})
                .get(arg_name, {})
                .get("type", None)
            )
            return arg_type == "string"
        logger.debug("No tool named '%s'.", tool_name)
        return False

    @staticmethod
    def _tools_enabled(request: ChatCompletionRequest) -> bool:
        """Return whether tool parsing should be applied for this request."""
        try:
            tools = getattr(request, "tools", None)
            tool_choice = getattr(request, "tool_choice", None)
            return bool(tools) and tool_choice != "none"
        except Exception:
            logger.exception("Failed to determine if tools are enabled.")
            return False

    def adjust_request(
        self, request: ChatCompletionRequest | ResponsesRequest
    ) -> ChatCompletionRequest | ResponsesRequest:
        """Adjust request parameters for tool call token handling."""
        request = super().adjust_request(request)
        if request.tools and request.tool_choice != "none":
            # Ensure tool call tokens (<tool_call>, </tool_call>) are not skipped
            # during decoding. Even though they are not marked as special tokens,
            # setting skip_special_tokens=False ensures proper handling in
            # transformers 5.x where decoding behavior may have changed.
            request.skip_special_tokens = False
        return request

    def extract_tool_calls(
        self,
        model_output: str,
        request: ChatCompletionRequest,
    ) -> ExtractedToolCallInformation:
        matched_tool_calls = self.func_call_regex.findall(model_output)
        logger.debug("model_output: %s", model_output)
        try:
            tool_calls: list[ToolCall] = []
            for match in matched_tool_calls:
                tc_detail = self.func_detail_regex.search(match)
                if not tc_detail:
                    logger.warning(
                        "Failed to parse tool call details from: %s",
                        match,
                    )
                    continue
                tc_name = tc_detail.group(1).strip()
                tc_args = tc_detail.group(2)
                pairs = self.func_arg_regex.findall(tc_args) if tc_args else []
                arg_dct: dict[str, Any] = {}
                for key, value in pairs:
                    arg_key = key.strip()
                    arg_val = value.strip()
                    if not self._is_string_type(tc_name, arg_key, request.tools):
                        arg_val = self._deserialize(arg_val)
                    logger.debug("arg_key = %s, arg_val = %s", arg_key, arg_val)
                    arg_dct[arg_key] = arg_val
                tool_calls.append(
                    ToolCall(
                        type="function",
                        function=FunctionCall(
                            name=tc_name,
                            arguments=json.dumps(arg_dct, ensure_ascii=False),
                        ),
                    )
                )
        except Exception:
            logger.exception("Failed to extract tool call spec")
            return ExtractedToolCallInformation(
                tools_called=False, tool_calls=[], content=model_output
            )
        else:
            if len(tool_calls) > 0:
                content: str | None = model_output[
                    : model_output.find(self.tool_calls_start_token)
                ]
                # Normalize empty/whitespace-only content to None
                if not content or not content.strip():
                    content = None
                return ExtractedToolCallInformation(
                    tools_called=True, tool_calls=tool_calls, content=content
                )
            return ExtractedToolCallInformation(
                tools_called=False, tool_calls=[], content=model_output
            )

    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:
        if not self._tools_enabled(request):
            return DeltaMessage(content=delta_text) if delta_text else None

        self._buffer += delta_text

        pending_deltas: dict[int, DeltaToolCall] = {}
        content: str | None = None

        while True:
            if not self._in_tool_call:
                start_idx = self._buffer.find(self.tool_call_start_token)
                if start_idx == -1:
                    # Check for partial start token at end of buffer
                    for i in range(1, len(self.tool_call_start_token)):
                        if self._buffer.endswith(self.tool_call_start_token[:i]):
                            out = self._buffer[:-i]
                            self._buffer = self._buffer[-i:]
                            if out:
                                content = (content or "") + out
                            break
                    else:
                        out = self._buffer
                        self._buffer = ""
                        if out:
                            content = (content or "") + out
                    break

                if start_idx > 0:
                    content = (content or "") + self._buffer[:start_idx]
                    self._buffer = self._buffer[start_idx:]

                self._buffer = self._buffer[len(self.tool_call_start_token) :]
                self._begin_tool_call()
                continue

            # Parse tool name first
            if not self.current_tool_name_sent:
                nl = self._buffer.find("\n")
                ak = self._buffer.find(self.arg_key_start)
                end = self._buffer.find(self.tool_call_end_token)
                candidates = [i for i in [nl, ak, end] if i != -1]
                if not candidates:
                    break
                cut = min(candidates)
                tool_name = self._buffer[:cut].strip()
                if tool_name == "" and cut == end:
                    # Handle empty tool call like `<tool_call></tool_call>`.
                    # Consume the tokens and reset state to avoid infinite loop.
                    self._buffer = self._buffer[end + len(self.tool_call_end_token) :]
                    self._finish_tool_call()
                    self._revert_last_tool_call_state()
                    continue

                if cut == nl:
                    self._buffer = self._buffer[nl + 1 :]
                else:
                    self._buffer = self._buffer[cut:]

                self._current_tool_name = tool_name
                self.current_tool_name_sent = True
                self._update_tool_name(pending_deltas, tool_name)
                continue

            assert self._current_tool_name is not None

            # Handle incremental string value streaming
            if self._streaming_string_value:
                val_end = self._buffer.find(self.arg_val_end)
                if val_end != -1:
                    raw_content = self._buffer[:val_end]
                    self._buffer = self._buffer[val_end + len(self.arg_val_end) :]
                    self._streaming_string_value = False
                    self._pending_key = None

                    escaped = self._json_escape_string_content(raw_content)
                    frag = escaped + '"'
                    self.streamed_args_for_tool[self.current_tool_id] += frag
                    self._update_tool_args(pending_deltas, frag)
                    continue

                # Check for partial </arg_value> at end
                safe_len = len(self._buffer)
                for i in range(1, len(self.arg_val_end)):
                    if self._buffer.endswith(self.arg_val_end[:i]):
                        safe_len = len(self._buffer) - i
                        break

                if safe_len > 0:
                    to_emit = self._buffer[:safe_len]
                    self._buffer = self._buffer[safe_len:]
                    escaped = self._json_escape_string_content(to_emit)
                    if escaped:
                        self.streamed_args_for_tool[self.current_tool_id] += escaped
                        self._update_tool_args(pending_deltas, escaped)
                break

            # If we have a pending key, parse its value
            if self._pending_key is not None:
                val_pos = self._buffer.find(self.arg_val_start)
                if val_pos == -1:
                    break
                if val_pos > 0:
                    self._buffer = self._buffer[val_pos:]

                key = (self._pending_key or "").strip()

                is_string = self._is_string_type(
                    self._current_tool_name, key, request.tools
                )

                if is_string:
                    # String type: stream incrementally
                    self._buffer = self._buffer[len(self.arg_val_start) :]

                    if key in self._seen_keys[self.current_tool_id]:
                        self._pending_key = None
                        continue

                    self._seen_keys[self.current_tool_id].add(key)
                    key_json = json.dumps(key, ensure_ascii=False)

                    if not self._args_started[self.current_tool_id]:
                        frag = "{" + key_json + ': "'
                        self._args_started[self.current_tool_id] = True
                    else:
                        frag = ", " + key_json + ': "'

                    self.streamed_args_for_tool[self.current_tool_id] += frag
                    self._streaming_string_value = True
                    self._update_tool_args(pending_deltas, frag)
                    continue

                # Non-string type: wait for complete value
                val_end = self._buffer.find(self.arg_val_end)
                if val_end == -1:
                    break

                raw_val = self._buffer[len(self.arg_val_start) : val_end].strip()
                self._buffer = self._buffer[val_end + len(self.arg_val_end) :]
                self._pending_key = None

                frag_or_none = self._append_arg_fragment(key=key, raw_val=raw_val)
                if frag_or_none:
                    self._update_tool_args(pending_deltas, frag_or_none)
                continue

            # Parse next arg or close
            end_pos = self._buffer.find(self.tool_call_end_token)
            key_pos = self._buffer.find(self.arg_key_start)
            if end_pos != -1 and (key_pos == -1 or end_pos < key_pos):
                self._buffer = self._buffer[end_pos + len(self.tool_call_end_token) :]
                frag_or_none = self._close_args_if_needed()
                # Finalize prev_tool_call_arr with complete parsed arguments
                if self._current_tool_name:
                    try:
                        full_args_str = self.streamed_args_for_tool[
                            self.current_tool_id
                        ]
                        args_dict = json.loads(full_args_str)
                        self.prev_tool_call_arr[self.current_tool_id] = {
                            "name": self._current_tool_name,
                            "arguments": args_dict,
                        }
                    except (json.JSONDecodeError, IndexError) as e:
                        logger.warning(
                            "Failed to finalize tool call state for tool %d: %s",
                            self.current_tool_id,
                            e,
                        )
                self._finish_tool_call()
                if frag_or_none:
                    self._update_tool_args(pending_deltas, frag_or_none)
                continue

            if key_pos == -1:
                break
            if key_pos > 0:
                self._buffer = self._buffer[key_pos:]
            key_end = self._buffer.find(self.arg_key_end)
            if key_end == -1:
                break
            key = self._buffer[len(self.arg_key_start) : key_end]
            self._buffer = self._buffer[key_end + len(self.arg_key_end) :]
            self._pending_key = key
            continue

        tool_calls = list(pending_deltas.values())
        if content is None and len(tool_calls) == 0:
            if request.logprobs:
                return DeltaMessage(content="")
            return None
        return DeltaMessage(content=content, tool_calls=tool_calls)

    def _ensure_tool_state(self) -> None:
        while len(self._tool_call_ids) <= self.current_tool_id:
            self._tool_call_ids.append(
                make_tool_call_id(id_type="random", func_name=None, idx=None)
            )
        while len(self.streamed_args_for_tool) <= self.current_tool_id:
            self.streamed_args_for_tool.append("")
        while len(self.prev_tool_call_arr) <= self.current_tool_id:
            self.prev_tool_call_arr.append({})
        while len(self._args_started) <= self.current_tool_id:
            self._args_started.append(False)
        while len(self._args_closed) <= self.current_tool_id:
            self._args_closed.append(False)
        while len(self._seen_keys) <= self.current_tool_id:
            self._seen_keys.append(set())

    def _begin_tool_call(self) -> None:
        if self.current_tool_id == -1:
            self.current_tool_id = 0
        else:
            self.current_tool_id += 1
        self._ensure_tool_state()
        self.current_tool_name_sent = False
        self._current_tool_name = None
        self._pending_key = None
        self._streaming_string_value = False
        self._in_tool_call = True

    def _finish_tool_call(self) -> None:
        self._in_tool_call = False
        self._current_tool_name = None
        self._pending_key = None
        self._streaming_string_value = False

    def _revert_last_tool_call_state(self) -> None:
        """Revert the state allocation for the last tool call."""
        if self.current_tool_id < 0:
            return
        self._tool_call_ids.pop()
        self.streamed_args_for_tool.pop()
        self.prev_tool_call_arr.pop()
        self._args_started.pop()
        self._args_closed.pop()
        self._seen_keys.pop()
        self.current_tool_id -= 1

    def _get_or_create_delta(self, pending: dict[int, DeltaToolCall]) -> DeltaToolCall:
        idx = self.current_tool_id
        if idx not in pending:
            pending[idx] = DeltaToolCall(
                index=idx,
                function=DeltaFunctionCall(),
            )
        delta = pending[idx]
        assert delta.function is not None
        return delta

    def _update_tool_name(
        self, pending: dict[int, DeltaToolCall], tool_name: str
    ) -> None:
        self.prev_tool_call_arr[self.current_tool_id] = {
            "name": self._current_tool_name,
            "arguments": {},
        }
        delta = self._get_or_create_delta(pending)
        delta.id = self._tool_call_ids[self.current_tool_id]
        delta.type = "function"
        assert delta.function is not None
        delta.function.name = tool_name
        if delta.function.arguments is None:
            delta.function.arguments = ""

    @staticmethod
    def _complete_json_prefix(
        json_prefix: str,
        allowed_partial_types: Allow,
    ) -> dict | None:
        """Complete a partial JSON prefix into a valid JSON object.

        Returns (formatted_prefix, parsed_dict) or None on failure.

        Note: ``partial_json_parser`` strips trailing whitespace before
        parsing (``complete.py:20``), which means the returned slice is
        shorter than ``json_prefix`` when it has trailing whitespace.
        Since the parser controls the construction of the json_prefix value,
        this code relies on it being a valid prefix and we only use the fix for
        the completion of the JSON object.
        """
        try:
            _, partial_str_completion = partial_json_parser.core.complete.fix(
                json_prefix,
                allowed_partial_types,
            )
            return json.loads(json_prefix + partial_str_completion)
        except Exception:
            return None

    def _update_tool_args(
        self, pending: dict[int, DeltaToolCall], fragment: str
    ) -> None:
        result = self._complete_json_prefix(
            self.streamed_args_for_tool[self.current_tool_id],
            Allow.ALL,
        )
        if result is not None:
            self.prev_tool_call_arr[self.current_tool_id]["arguments"] = result
        delta = self._get_or_create_delta(pending)
        assert delta.function is not None
        if delta.function.arguments is None:
            delta.function.arguments = ""
        delta.function.arguments += fragment

    def _append_arg_fragment(
        self,
        *,
        key: str,
        raw_val: str,
    ) -> str | None:
        key = key.strip()
        if not key:
            return None
        if key in self._seen_keys[self.current_tool_id]:
            return None

        # This function is only called for non-string types (already checked
        # by _is_string_type in the caller), so we always deserialize.
        val_obj: Any = self._deserialize(raw_val)

        key_json = json.dumps(key, ensure_ascii=False)
        val_json = json.dumps(val_obj, ensure_ascii=False)

        if not self._args_started[self.current_tool_id]:
            fragment = "{" + key_json + ": " + val_json
            self._args_started[self.current_tool_id] = True
        else:
            fragment = ", " + key_json + ": " + val_json

        self._seen_keys[self.current_tool_id].add(key)
        self.streamed_args_for_tool[self.current_tool_id] += fragment
        return fragment

    def _close_args_if_needed(self) -> str | None:
        if self._args_closed[self.current_tool_id]:
            return None
        self._args_closed[self.current_tool_id] = True
        if not self._args_started[self.current_tool_id]:
            fragment = "{}"
            self.streamed_args_for_tool[self.current_tool_id] = fragment
        else:
            fragment = "}"
            self.streamed_args_for_tool[self.current_tool_id] += fragment
        return fragment

_complete_json_prefix staticmethod

_complete_json_prefix(
    json_prefix: str, allowed_partial_types: Allow
) -> dict | None

Complete a partial JSON prefix into a valid JSON object.

Returns (formatted_prefix, parsed_dict) or None on failure.

Note: partial_json_parser strips trailing whitespace before parsing (complete.py:20), which means the returned slice is shorter than json_prefix when it has trailing whitespace. Since the parser controls the construction of the json_prefix value, this code relies on it being a valid prefix and we only use the fix for the completion of the JSON object.

Source code in vllm/tool_parsers/poolside_v1_tool_parser.py
@staticmethod
def _complete_json_prefix(
    json_prefix: str,
    allowed_partial_types: Allow,
) -> dict | None:
    """Complete a partial JSON prefix into a valid JSON object.

    Returns (formatted_prefix, parsed_dict) or None on failure.

    Note: ``partial_json_parser`` strips trailing whitespace before
    parsing (``complete.py:20``), which means the returned slice is
    shorter than ``json_prefix`` when it has trailing whitespace.
    Since the parser controls the construction of the json_prefix value,
    this code relies on it being a valid prefix and we only use the fix for
    the completion of the JSON object.
    """
    try:
        _, partial_str_completion = partial_json_parser.core.complete.fix(
            json_prefix,
            allowed_partial_types,
        )
        return json.loads(json_prefix + partial_str_completion)
    except Exception:
        return None

_json_escape_string_content staticmethod

_json_escape_string_content(s: str) -> str

JSON-escape string content for incremental streaming.

This escapes the content that goes INSIDE a JSON string (between quotes), not including the surrounding quotes themselves.

Source code in vllm/tool_parsers/poolside_v1_tool_parser.py
@staticmethod
def _json_escape_string_content(s: str) -> str:
    """JSON-escape string content for incremental streaming.

    This escapes the content that goes INSIDE a JSON string (between quotes),
    not including the surrounding quotes themselves.
    """
    if not s:
        return ""
    return json.dumps(s, ensure_ascii=False)[1:-1]

_revert_last_tool_call_state

_revert_last_tool_call_state() -> None

Revert the state allocation for the last tool call.

Source code in vllm/tool_parsers/poolside_v1_tool_parser.py
def _revert_last_tool_call_state(self) -> None:
    """Revert the state allocation for the last tool call."""
    if self.current_tool_id < 0:
        return
    self._tool_call_ids.pop()
    self.streamed_args_for_tool.pop()
    self.prev_tool_call_arr.pop()
    self._args_started.pop()
    self._args_closed.pop()
    self._seen_keys.pop()
    self.current_tool_id -= 1

_tools_enabled staticmethod

_tools_enabled(request: ChatCompletionRequest) -> bool

Return whether tool parsing should be applied for this request.

Source code in vllm/tool_parsers/poolside_v1_tool_parser.py
@staticmethod
def _tools_enabled(request: ChatCompletionRequest) -> bool:
    """Return whether tool parsing should be applied for this request."""
    try:
        tools = getattr(request, "tools", None)
        tool_choice = getattr(request, "tool_choice", None)
        return bool(tools) and tool_choice != "none"
    except Exception:
        logger.exception("Failed to determine if tools are enabled.")
        return False

adjust_request

Adjust request parameters for tool call token handling.

Source code in vllm/tool_parsers/poolside_v1_tool_parser.py
def adjust_request(
    self, request: ChatCompletionRequest | ResponsesRequest
) -> ChatCompletionRequest | ResponsesRequest:
    """Adjust request parameters for tool call token handling."""
    request = super().adjust_request(request)
    if request.tools and request.tool_choice != "none":
        # Ensure tool call tokens (<tool_call>, </tool_call>) are not skipped
        # during decoding. Even though they are not marked as special tokens,
        # setting skip_special_tokens=False ensures proper handling in
        # transformers 5.x where decoding behavior may have changed.
        request.skip_special_tokens = False
    return request