Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+intelligence_score_vs_output_tokens.png filter=lfs diff=lfs merge=lfs -text
+*.json filter=lfs diff=lfs merge=lfs -text
+*.txt filter=lfs diff=lfs merge=lfs -text
+tokenizer* filter=lfs diff=lfs merge=lfs -text
+*.tokenizer filter=lfs diff=lfs merge=lfs -text

model-00003-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac592467a2391b507296d568a87c6f02d41a1a3473f3bc6705fdafe91a7e84d9
+size 4302382136

model-00005-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d4cd762a7f2575c8a8fb04a68556e2311350567c9cb015fd94af7d37e5ef5c7d
+size 4302381912

model-00008-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8659a4124c5001e4f594fc53586c148c3210c8c080efcd0ebf420df1d18a3d9
+size 4302382112

model-00015-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ded3c735c606e9f1bc067f7782b63d701577e59a338b08e3fc49c8fb603be3f8
+size 4302347992

model-00018-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab55b759fa61688e5cdbfc0e76426e29eee60cb7dc4098a505268bd8508105c0
+size 4302347800

model-00025-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c4f8ea8cfd59a26c06a4118e98f5519d25f9471c7f2dc09534ec0352cb4efcf
+size 4302382720

model-00027-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9d5999b2943768dcc67cad1f8411d8f1e1aeb55c429e513d7789a5cad3152be
+size 4302382448

model-00036-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80dfc11f732d16b60d7e150ec6bffe84cf6373f9963383fc1da6ec1491a04b72
+size 4302382672

model-00043-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c31b7cb0d0e6761fec2ce693b99059802b055a51cac5d7e0485d214517af4cdf
+size 4302348184

model-00045-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d205fd835afec706efad9b7349b92bc098e1e21d7c4c283ec7aad45423d381c8
+size 4302348584

model-00046-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf167f9e7c7b2c5a91d6af1a2a4b3d837693c751045a767b7754459ae0bbb15c
+size 4302382264

model-00049-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49f78faa1b3a5dd74244c87806000e8cc53a75c5d904f40a7e84eecfe0c43882
+size 4302382448

model-00059-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82cb401b037eb3a5d1fa2ddff2f6aada163be92a5ce4263efc29ea1b0abd0dd9
+size 4302348568

model-00076-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3733153525cf7775f6f1ecd5c4a7285f742a74ec9d9390baf06bb0ddc28fa17
+size 4302382248

model-00083-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c25a777958d3a868ab6b57b6e5d2d9e040a0fa95d8a67c095a21e03a7e1a04b4
+size 4302382720

model-00084-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0df8afe9fc329bd69bd0f0349240fe622c8f2de176e680699894e36f0f21c78f
+size 4302348376

model-00086-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f63d1bb77a82c1f0c2f84f2fdb5c5a81ee8eead300feeb333c3bb52f2f007c9
+size 4302382720

model-00088-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5740f8e77b3f86fd8c686d5eda884eea8c637a1a4937046f44f988a27543041e
+size 4302382656

model-00089-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3137b2b82b3d50c80e48e187e6c743d735124bd55a9ecf2b6c5a4f008a9f73a1
+size 4302348584

model-00090-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e7368bc72f3f331d5bb6ab9ca023b42a4423eb21266e21e8d30f1cd70b9cd4f
+size 4302382264

model-00092-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c512f6f8a69f4528f503841f9982d39c044a5bd077939505e5b763de435502f9
+size 4302348392

model-00094-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3bfb9aef946581c145c358c416e687a4271c49848daa7edae10ddd1575b325a
+size 4302382720

model-00096-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5951c6ff635b0b2b1f4a3bce3158f71ce16b27f51a6e75f2b91296deacb52b9c
+size 4302382640

model-00105-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fbbcd10373f9a2dfc33a650a117cf000e7117fa863bbf0835d57ca47ef1c8770
+size 4302382720

model-00107-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9acd49a549e874ae346c940dc939144670da931d6f3d3ca69f1de7ae588399d
+size 4302382472

model-00108-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5378084e6769b6a496be211d5808858149956aca7893d5e928c2772f5f9e476
+size 4302382720

model-00109-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:820b730ab4882f28492381fd40c0591ac68e276e69450d5e3ed553496dc2a349
+size 4302348184

model-00110-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e37571416f241afcc413442a11ce8896a1ee63aab13b065227c0a04099cb2c15
+size 4302382656

model-00111-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:605b39268636cae3bb925ef22c34c720840bc6b4ee33366fc5c98e8d8d4a5d83
+size 4302348584

model-00112-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7152fa20f96fa500166c52abdebdfd00c19871d81f235166522d5169e7263ce
+size 4302382264

model-00113-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ea651e865707be3259305153f4e7a06c6f4c6b4d9989bb6b97f92039d7714d0
+size 4302382720

model-00115-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35f808fc0eb23a376220e074d4463ade7f685eb174a528aac37f56b90a9c241f
+size 4302382448

model-00116-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef8f050ca1cc24dca24f789af6e9c81cb7da3e33163a9dd753bf9ac983799387
+size 4302382720

model-00119-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94d7b0e6a2393122e0cf50556d05ed050b2016f71cf09a8265c89361feff9e9b
+size 4302348600

model-00123-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:519568466224076a0751d3bfe99f1e1c2e63b063df7042412777f74bb78fc6e5
+size 4302315568

model-00125-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4a0dd8ee862041575bd6747e29a31decc36ee61e06286ee9fde205009391bd1
+size 4302348568

model-00126-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2123b22cd5f9deca6155053114f874e45d74033f66b0e98878d91abaec10526f
+size 4302382280

model-00133-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f488efd2ab0864520342e8da6696b711a2fc67440e5093c9cb6856b0e24a1f0
+size 4302348584

model-00144-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e1d10df8855972ca9c778f59a13389ffe22b617b80b443859aa232cd2bed0be
+size 4302348568

model-00154-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5035c3fc7b727c6087ce3eee4de9a2979d821b7ccfbf61c793bb2ea13634f2dd
+size 4302382720

model-00156-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7b340d7dfee6e91cc40a124255a95b459be3076d533812ce2bf2cb12a569871
+size 4302382448

model-00161-of-000163.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8410fc54d2cca5a033964a922bc7f7aa954e7d549f5843606a4ab6ce1e57ccb
+size 4302382080

tool_parser_vllm.py

@@ -0,0 +1,583 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# ruff: noqa
+import json
+from collections.abc import Sequence
+from enum import Enum
+from typing import Any, Union
+
+import partial_json_parser
+import regex as re
+from partial_json_parser.core.options import Allow
+
+from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
+                                              DeltaFunctionCall, DeltaMessage,
+                                              DeltaToolCall,
+                                              ExtractedToolCallInformation,
+                                              FunctionCall, ToolCall)
+from vllm.entrypoints.openai.tool_parsers.abstract_tool_parser import (
+    ToolParser, ToolParserManager)
+from vllm.logger import init_logger
+from vllm.transformers_utils.tokenizer import AnyTokenizer
+from vllm.utils import random_uuid
+
+logger = init_logger(__name__)
+
+
+class ParsedStructure(Enum):
+    CONTENT = 1
+    REASONING_CONTENT = 2
+    TOOL_CALL = 3
+    TOOL_CALL_DELIMITER = 4
+    TOOL_CALL_START_TAG = 5
+    TOOL_CALL_END_TAG = 6
+
+
+@ToolParserManager.register_module("tng_r1t2")
+class TngR1T2ToolParser(ToolParser):
+    """Tool Parser for models like tngtech/DeepSeek-TNG-R1T2-Chimera,
+    It is compatible with hermes tool call templates
+    but does not require <tool_call> and </tool_call>
+    to be single tokens in the vocabulary;
+    instead only the string representation of the model output
+    is parsed, making this tool call parser robust and versatile."""
+
+    def __init__(self, tokenizer: AnyTokenizer):
+        super().__init__(tokenizer)
+
+        # For backward compatibility with serving code
+        self.prev_tool_call_arr: list[dict] = []
+        self.streamed_args_for_tool: list[str] = []
+
+        self.think_tag_pattern = r"(<think>[\s\S]*?</think>)"
+        self.think_start_tag = "<think>"
+        self.think_end_tag = "</think>"
+
+        self.tool_call_tag_pattern = r"<tool_call>([\s\S]*?)</tool_call>"
+        self.tool_call_start_tag = "<tool_call>"
+        self.tool_call_end_tag = "</tool_call>"
+
+        # Define streaming state type to be initialized later
+        self.streaming_state: dict[str, Any] = {
+            "streamed_tool_calls": [],
+            "buffer": "",
+            "parsed_structure": ParsedStructure.CONTENT,
+        }
+
+    def extract_tool_call_from_nonthink_output(
+            self, raw_text: str) -> tuple[str, list[dict]]:
+        parts = re.split(self.tool_call_tag_pattern, raw_text)
+        content = ""
+        tool_calls: list[dict] = []
+        for i, part in enumerate(parts):
+            is_potential_tool_call = i % 2 == 1
+            if is_potential_tool_call:
+                try:
+                    more_tool_calls = json.loads(part)
+                    if isinstance(more_tool_calls, list):
+                        tool_calls.extend(more_tool_calls)
+                    else:
+                        tool_calls.extend([more_tool_calls])
+                except json.JSONDecodeError:
+                    logger.warning("Invalid tool call json "
+                                   "-> parse as text content")
+                    content += part
+                    continue
+            else:
+                content += part
+        return content, tool_calls
+
+    def extract_tool_calls(
+        self,
+        model_output: str,
+        request: ChatCompletionRequest,
+    ) -> ExtractedToolCallInformation:
+        """
+        Extract tool calls from a complete model output.
+        """
+        # split at think traces -> those will not be parsed for tool calls
+        think_parts = re.split(self.think_tag_pattern, model_output)
+        content = ""
+        tool_calls = []
+        for i, part in enumerate(think_parts):
+            parse_output = i % 2 == 0
+            if parse_output:
+                more_content, more_tool_calls = (
+                    self.extract_tool_call_from_nonthink_output(part))
+                content += more_content
+                tool_calls += more_tool_calls
+            else:
+                content += part
+
+        if not tool_calls:
+            return ExtractedToolCallInformation(
+                tools_called=False,
+                tool_calls=[],
+                content=content,
+            )
+
+        tool_call_objs: list[ToolCall] = []
+
+        for idx, call in enumerate(tool_calls):
+            if (not isinstance(call, dict) or "name" not in call
+                    or "arguments" not in call):
+                logger.warning("Invalid tool call format, ignore.")
+                continue
+
+            tool_call = ToolCall(
+                id=f"call_{idx}_{random_uuid()}",
+                type="function",
+                function=FunctionCall(
+                    name=call["name"],
+                    arguments=(json.dumps(call["arguments"]) if isinstance(
+                        call["arguments"], dict) else call["arguments"]),
+                ),
+            )
+            tool_call_objs.append(tool_call)
+
+        return ExtractedToolCallInformation(
+            tools_called=len(tool_call_objs) > 0,
+            tool_calls=tool_call_objs,
+            content=content,
+        )
+
+    def _parse_think_trace(self, raw_text: str) -> tuple[str, bool, str]:
+        """
+        Returns: (unambiguous_text_content, found_think_end, rest_string)
+        """
+        # Either a complete think_end_tag can be somewhere in raw_text
+        think_end_pos = raw_text.find(self.think_end_tag)
+        if think_end_pos >= 0:
+            # in contrast to tool_call_start_tags, </think> remains part of content
+            think_end_tag_end_pos = think_end_pos + len(self.think_end_tag)
+            return (raw_text[:think_end_tag_end_pos], True,
+                    raw_text[think_end_tag_end_pos:])
+
+        # or the end of raw_text can be continued to a complete think_end_tag
+        think_end_pos = (
+            len(raw_text) -
+            self._ends_with_partial_token(raw_text, self.think_end_tag))
+        return raw_text[:think_end_pos], False, raw_text[think_end_pos:]
+
+    def _parse_unambiguous_text_content(
+            self, raw_text: str) -> tuple[str, Union[str, None], str]:
+        """
+        Returns: (unambiguous_text_content, interrupting_tag, rest_string)
+        """
+        # Either a complete tool_call_start_tag or think_start can be somewhere in raw_text
+        search_tags = [self.think_start_tag, self.tool_call_start_tag]
+        tag_positions = [(tag, pos) for tag in search_tags
+                         if (pos := raw_text.find(tag)) >= 0]
+        tag_positions.sort(key=lambda tag_and_pos: tag_and_pos[1])
+        if len(tag_positions) > 0:
+            first_tag, tag_pos = tag_positions[0]
+            return raw_text[:tag_pos], first_tag, raw_text[tag_pos:]
+
+        # or the end of raw_text can be continued to a complete tag
+        tag_positions = [
+            (tag, len(raw_text) - self._ends_with_partial_token(raw_text, tag))
+            for tag in search_tags
+        ]
+        tag_positions.sort(key=lambda tag_and_pos: tag_and_pos[1])
+        first_tag, tag_pos = tag_positions[0]
+        if tag_pos < len(raw_text):
+            return raw_text[:tag_pos], None, raw_text[tag_pos:]
+        return raw_text, None, ""
+
+    def _parse_tool_call_start_tag(self, raw_text: str) -> tuple[bool, str]:
+        """
+        Removes tool_call_start_tag from the beginning of raw_text,
+        and an optional "[", and leading whitespace.
+        Returns: (found_complete_tool_call_start_tag, rest_string)
+        """
+        if not raw_text.startswith(self.tool_call_start_tag):
+            return False, raw_text
+        rest = raw_text[len(self.tool_call_start_tag):].lstrip()
+        if rest.startswith("["):
+            rest = rest[1:].lstrip()
+        return True, rest
+
+    def _parse_tool_call_end_tag(
+            self, raw_text: str) -> tuple[Union[bool, None], str]:
+        """
+        Removes tool_call_end_tag from the beginning of raw_text,
+        and an optional "]" before it, and leading whitespace.
+        Returns: tuple
+            found_complete_tool_call_end_tag (or None if not decidable yet)
+            rest_string
+        """
+        # remove optional whitespace and closing ] bracket from json list notation
+        rest = raw_text.lstrip()
+        if rest.startswith("]"):
+            rest = rest[1:].lstrip()
+
+        if rest.startswith(self.tool_call_end_tag):
+            # found a complete tool call end tag
+            return True, rest[len(self.tool_call_end_tag):]
+        if (len(rest) >= len(self.tool_call_end_tag)
+                or rest != self.tool_call_end_tag[:len(rest)]):
+            # evidence that rest_string does not start with a tool call end tag
+            return False, raw_text
+        # incomplete tool call end tag, can not be decided yet
+        return None, raw_text
+
+    def _extract_arguments_from_partial_tool_call(
+            self, raw_text: str) -> Union[str, None]:
+        """
+        Extracts the raw text of the "arguments" field of a complete
+        or partial tool call.
+        Args:
+            raw_text: tool call raw text,
+                      e.g `{"name": "my_tool", "arguments": {"firstarg": "some`
+
+        Returns:
+            raw text of the "arguments" field, which is not valid JSON
+            unless the tool call is complete,
+            e.g. `{"firstarg": "some` for the example raw_text above
+        """
+        # assumptions:
+        # - "arguments" is always an object
+        # - there is no other field of type object in the function call
+        # - `raw_text` contains first "name", then "arguments" (otherwise,
+        #   we'd have to find the end of "arguments" before returning its
+        #   raw text value)
+
+        # typically, at position 0, but there might be leading whitespace
+        tool_call_start_pos = raw_text.find("{")
+        assert raw_text[:tool_call_start_pos].strip() == ""
+        arguments_start_pos = raw_text.find("{", tool_call_start_pos + 1)
+        if arguments_start_pos < 0:
+            return None
+        arguments_raw_text = raw_text[arguments_start_pos:]
+        return arguments_raw_text
+
+    def _parse_complete_tool_call(
+            self, raw_text: str) -> tuple[Union[dict, None], str]:
+        """
+        Returns: tuple
+            parsed tool call if complete, None otherwise
+            rest_string that needs to be parsed again or may contain
+                        a partial tool call
+        """
+        # raw_text must start without whitespace for correct parsing
+        obj, end_pos = self.extract_complete_json_dict(raw_text)
+        if obj is None:
+            return None, raw_text
+        tool_call_raw_text = raw_text[:end_pos]
+        # `tool_call_raw_text` is something like:
+        #   '{"name": "tool-name", "arguments": {...xyz...} }'
+        # we want to extract `{...xyz...}`,
+        # but `extract_arguments_from_partial_tool_call` would return
+        # everything after the second '{', i.e. `{...xyz...} }`
+        arguments_raw_text = self._extract_arguments_from_partial_tool_call(
+            tool_call_raw_text.removesuffix("}").rstrip())
+        tool_call = {
+            "name": obj.get("name"),
+            "arguments": obj.get("arguments"),
+            "arguments_raw_text": arguments_raw_text,
+            "is_complete": True,
+        }
+        return tool_call, raw_text[end_pos:]
+
+    def _parse_partial_tool_call(self, raw_text: str) -> Union[dict, None]:
+        # raw_text must start without whitespace for correct parsing
+        obj = partial_json_parser.loads(raw_text, Allow.ALL)
+        arguments_raw_text = (
+            self._extract_arguments_from_partial_tool_call(raw_text))
+        tool_call = {
+            "name": obj.get("name"),
+            "arguments": obj.get("arguments"),
+            "arguments_raw_text": arguments_raw_text,
+            "is_complete": False,
+        }
+        return tool_call
+
+    def _parse_tool_call(
+            self,
+            raw_text: str) -> tuple[Union[bool, None], Union[dict, None], str]:
+        # remove optional whitespace and closing ] bracket
+        # from json list notation
+        rest = raw_text.lstrip()
+        if rest == "":
+            # no json has been received yet
+            # -> can't tell if this will be a valid tool call
+            return None, None, raw_text
+        if not rest.startswith("{"):
+            # can't be a tool call json
+            return False, None, raw_text
+
+        tool_call, rest = self._parse_complete_tool_call(rest)
+        if tool_call:
+            return True, tool_call, rest
+
+        try:
+            tool_call = self._parse_partial_tool_call(rest)
+            # need to re-parse partial tool call later again
+            # -> return None, not True
+            return None, tool_call, rest
+        except json.JSONDecodeError:
+            # invalid json -> neither complete nor partial tool call
+            return False, None, rest
+
+    def _parse_tool_call_delimiter(
+            self, raw_text: str) -> tuple[Union[bool, None], str]:
+        """
+        Returns: tuple
+            does raw_text start with tool call delimiter?
+                (None if undecidable/incomplete)
+            rest_string
+        """
+        rest = raw_text.lstrip()
+        if rest == "":
+            return None, raw_text
+        has_next_tool_call = rest.startswith(",")
+        if not has_next_tool_call:
+            return False, raw_text
+
+        rest = rest[1:].lstrip()
+        if rest == "":
+            return None, raw_text
+        has_next_tool_call = rest.startswith("{")
+        if not has_next_tool_call:
+            return False, raw_text
+        return True, rest
+
+    def _parse_all(
+        self, raw_text: str, start_mode: ParsedStructure
+    ) -> tuple[str, list[dict], str, ParsedStructure]:
+        if start_mode == ParsedStructure.REASONING_CONTENT:
+            content, found_closing_think, rest = self._parse_think_trace(
+                raw_text)
+            if found_closing_think:
+                more_content, tool_calls, rest, structure = self._parse_all(
+                    rest, start_mode=ParsedStructure.CONTENT)
+                return content + more_content, tool_calls, rest, structure
+            return content, [], rest, ParsedStructure.REASONING_CONTENT
+
+        elif start_mode == ParsedStructure.CONTENT:
+            content, interrupting_tag, rest = self._parse_unambiguous_text_content(
+                raw_text)
+
+            # rest might contain a tool call start tag or a think start tag
+            if interrupting_tag == self.tool_call_start_tag:
+                more_content, tool_calls, rest, structure = self._parse_all(
+                    rest, start_mode=ParsedStructure.TOOL_CALL_START_TAG)
+                return content + more_content, tool_calls, rest, structure
+            elif interrupting_tag == self.think_start_tag:
+                more_content, tool_calls, rest, structure = self._parse_all(
+                    rest, start_mode=ParsedStructure.REASONING_CONTENT)
+                return content + more_content, tool_calls, rest, structure
+            else:
+                return content, [], rest, ParsedStructure.CONTENT
+
+        elif start_mode == ParsedStructure.TOOL_CALL_START_TAG:
+            found_tool_call_start_tag, rest = self._parse_tool_call_start_tag(
+                raw_text)
+            if not found_tool_call_start_tag:
+                return "", [], raw_text, ParsedStructure.CONTENT
+            # we found a complete start tag, but we haven't seen the begin of a tool call json yet
+            content, tool_calls, rest, structure = self._parse_all(
+                rest, start_mode=ParsedStructure.TOOL_CALL)
+            if not content and not tool_calls:
+                # We haven't reached the opening "{" of the tool call yet.
+                # We might see a "[" before the "{", so let's process the start tag again next chunk.
+                return content, [], raw_text, ParsedStructure.CONTENT
+            return content, tool_calls, rest, structure
+
+        elif start_mode == ParsedStructure.TOOL_CALL:
+            found_tool_call, tool_call, rest = self._parse_tool_call(raw_text)
+            if found_tool_call is True:
+                tool_calls = [tool_call] if tool_call else []
+                content, more_tool_calls, rest, structure = self._parse_all(
+                    rest, start_mode=ParsedStructure.TOOL_CALL_DELIMITER)
+                return (content, tool_calls + more_tool_calls, rest, structure)
+            elif found_tool_call is None:
+                # partial tool call -> need to parse again with next chunk
+                tool_calls = ([tool_call] if tool_call is not None else [])
+                return "", tool_calls, rest, ParsedStructure.TOOL_CALL
+            else:
+                logger.warning(
+                    "Invalid tool call -> continue with parsing model output as text content"
+                )
+                return self._parse_all(raw_text,
+                                       start_mode=ParsedStructure.CONTENT)
+
+        elif start_mode == ParsedStructure.TOOL_CALL_DELIMITER:
+            found_tool_call_delimiter, rest = self._parse_tool_call_delimiter(
+                raw_text)
+            if found_tool_call_delimiter is True:
+                return self._parse_all(rest,
+                                       start_mode=ParsedStructure.TOOL_CALL)
+            elif found_tool_call_delimiter is None:
+                # could neither confirm nor deny that raw_text starts with a tool call delimiter
+                return "", [], rest, ParsedStructure.TOOL_CALL_DELIMITER
+            else:
+                return self._parse_all(
+                    raw_text, start_mode=ParsedStructure.TOOL_CALL_END_TAG)
+
+        elif start_mode == ParsedStructure.TOOL_CALL_END_TAG:
+            found_tool_call_end_tag, rest = self._parse_tool_call_end_tag(
+                raw_text)
+            if found_tool_call_end_tag is True:
+                return self._parse_all(rest,
+                                       start_mode=ParsedStructure.CONTENT)
+            elif found_tool_call_end_tag is None:
+                return "", [], rest, ParsedStructure.TOOL_CALL_END_TAG
+            else:
+                return self._parse_all(raw_text,
+                                       start_mode=ParsedStructure.CONTENT)
+
+        logger.warning(
+            f"Unknown tool call parser start_mode '{start_mode}'. Falling back to text content."
+        )
+        return self._parse_all(raw_text, start_mode=ParsedStructure.CONTENT)
+
+    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,
+    ) -> Union[DeltaMessage, None]:
+        """
+        Extract tool calls for streaming mode.
+        """
+        raw_text = self.streaming_state["buffer"] + delta_text
+        structure = self.streaming_state["parsed_structure"]
+
+        content, tool_calls, rest, new_structure = self._parse_all(
+            raw_text, start_mode=structure)
+        self.streaming_state["buffer"] = rest
+        self.streaming_state["parsed_structure"] = new_structure
+
+        already_streamed_tool_calls = (
+            self.streaming_state["streamed_tool_calls"])
+        already_streamed_complete_tool_calls = [
+            tool_call for tool_call in already_streamed_tool_calls
+            if tool_call["is_complete"]
+        ]
+        all_tool_calls = (already_streamed_complete_tool_calls +
+                          (tool_calls or []))
+        to_be_streamed_tool_calls = self._calculate_delta_tool_calls(
+            all_tool_calls, already_streamed_tool_calls)
+
+        if not content and not to_be_streamed_tool_calls:
+            return None
+        self.update_state_vars(all_tool_calls)
+        return DeltaMessage(content=content if content else None,
+                            tool_calls=to_be_streamed_tool_calls)
+
+    def _calculate_delta_tool_calls(
+            self, current_tool_calls: Union[list[dict], None],
+            already_streamed_tool_calls: list[dict]) -> list[DeltaToolCall]:
+        if not current_tool_calls:
+            return []
+
+        new_deltas = []
+        for tool_call_idx, partial_tool_call in enumerate(current_tool_calls):
+            if (partial_tool_call.get("name") is None
+                    or partial_tool_call.get("arguments") is None):
+                # do not stream arguments for an unknown tool name;
+                # and unless arguments appear in the partial json,
+                # it might be that "name" has not been received completely
+                # (assuming a template like `{"name": "mytool", "arguments": ...}`)
+                continue
+            partial_tool_call["tool_call_idx"] = tool_call_idx
+            partial_tool_call["arguments_raw_text"] = (
+                partial_tool_call.get('arguments_raw_text') or "")
+
+            if len(already_streamed_tool_calls) > tool_call_idx:
+                # parts of this tool_call_idx have already been streamed
+                already_streamed_tool_call = already_streamed_tool_calls[
+                    tool_call_idx]
+                delta_tool_call = self._delta_for_partial_tool_call(
+                    partial_tool_call, already_streamed_tool_call)
+                if delta_tool_call is not None:
+                    new_deltas.append(delta_tool_call)
+                already_streamed_tool_calls[tool_call_idx] = (
+                    already_streamed_tool_call | partial_tool_call)
+            else:
+                # no parts of this tool_call_idx have been streamed yet
+                tool_call = self._delta_for_new_tool_call(partial_tool_call)
+                new_deltas.append(tool_call)
+                already_streamed_tool_calls.append(partial_tool_call)
+
+        return new_deltas
+
+    def _delta_for_new_tool_call(self, tool_call_dict: dict) -> DeltaToolCall:
+        """constructs DeltaToolCall for new tool call,
+        with tool_call_id, name, and all arguments seen so far.
+        Updates tool_call dictionary with tool_call_id.
+        """
+        tool_call_idx = tool_call_dict["tool_call_idx"]
+        tool_call_dict[
+            "tool_call_id"] = f"call_{tool_call_idx}_{random_uuid()}"
+        tool_call_dict["arguments_raw_text"] = tool_call_dict.get(
+            'arguments_raw_text') or ""
+        delta_tool_call = DeltaToolCall(
+            index=tool_call_idx,
+            type="function",
+            id=tool_call_dict["tool_call_id"],
+            function=DeltaFunctionCall(
+                name=tool_call_dict.get("name"),
+                arguments=tool_call_dict["arguments_raw_text"]))
+        return delta_tool_call
+
+    def _delta_for_partial_tool_call(
+            self, new_tool_call: dict,
+            already_streamed_tool_call: dict) -> Union[DeltaToolCall, None]:
+        """Calculate delta for a tool call of which some parts have already been streamed."""
+        assert new_tool_call["name"] == already_streamed_tool_call["name"]
+        assert already_streamed_tool_call.get("tool_call_id")
+        if already_streamed_tool_call.get("is_complete"):
+            return None
+
+        to_be_streamed_arguments = (
+            new_tool_call["arguments_raw_text"].removeprefix(
+                already_streamed_tool_call["arguments_raw_text"]))
+        if not to_be_streamed_arguments:
+            return None
+
+        delta_tool_call = DeltaToolCall(
+            index=new_tool_call["tool_call_idx"],
+            type="function",
+            function=DeltaFunctionCall(arguments=to_be_streamed_arguments))
+        return delta_tool_call
+
+    def update_state_vars(self, all_tools: list[dict]) -> None:
+        # `tool_parser.streamed_args_for_tool` and
+        # `tool_parser.prev_tool_call_arr` are checked in serving_chat.py
+
+        # relevant is {"arguments": {...}}
+        self.prev_tool_call_arr = all_tools
+
+        # json-serialized argument
+        self.streamed_args_for_tool = [
+            tool_call.get("arguments_raw_text", "") for tool_call in all_tools
+        ]
+
+    @classmethod
+    def _ends_with_partial_token(cls, buffer: str, tag: str) -> int:
+        """
+        Check if buffer ends with a partial tag.
+        Return the length of the partial tag.
+        """
+        for i in range(1, min(len(buffer) + 1, len(tag))):
+            if tag.startswith(buffer[-i:]):
+                return i
+        return 0
+
+    @classmethod
+    def extract_complete_json_dict(cls, json_str: str):
+        try:
+            decoder = json.JSONDecoder()
+            obj, end_pos = decoder.raw_decode(
+                json_str)  # ignore any text after the end of the json object
+            if isinstance(obj, dict):
+                return obj, end_pos
+            return None, 0
+        except json.JSONDecodeError:
+            return None, 0