add Tool Parser (#3272)

* add tool-parser

* add tool-parser

* add tool parser

* add tool parser

* fix

* add offline

* add offline

* fix

* parsers:tool&reasoning

* 修改tool parser名称·

* update

* fix reasoning-parser

* add requirements

* fix finish reason

* fix

* fix reasoning-parser

* fix

* fix

* fix

* fix

* fix

---------

Co-authored-by: zhuzixuan <zhuzixuan@baidu.com>

fastdeploy/engine/args_utils.py

@@ -15,10 +15,10 @@
 """
 
 import json
+import os
 from dataclasses import asdict, dataclass
 from dataclasses import fields as dataclass_fields
 from typing import Any, Dict, List, Optional
-import os
 
 from fastdeploy.config import (
     CacheConfig,
@@ -94,6 +94,14 @@ class EngineArgs:
     """
     specifies the reasoning parser to use for extracting reasoning content from the model output
     """
+    tool_call_parser: str = None
+    """
+    specifies the tool call parser  to use for extracting tool call from the model output
+    """
+    tool_parser_plugin: str = None
+    """
+    tool parser plugin used to register user defined tool parsers
+    """
     enable_mm: bool = False
     """
     Flags to enable multi-modal model
@@ -434,6 +442,18 @@ class EngineArgs:
             help="Flag specifies the reasoning parser to use for extracting "
             "reasoning content from the model output",
         )
+        model_group.add_argument(
+            "--tool-call-parser",
+            type=str,
+            default=EngineArgs.tool_call_parser,
+            help="Flag specifies the tool call parser to use for extracting" "tool call from the model output",
+        )
+        model_group.add_argument(
+            "--tool-parser-plugin",
+            type=str,
+            default=EngineArgs.tool_parser_plugin,
+            help="tool parser plugin used to register user defined tool parsers",
+        )
         model_group.add_argument(
             "--speculative-config",
             type=json.loads,
@@ -885,10 +905,10 @@ class EngineArgs:
             if self.enable_chunked_prefill:
                 self.max_num_batched_tokens = 2048
             else:
-                if not int(os.getenv('ENABLE_V1_KVCACHE_SCHEDULER', '0')):
+                if not int(os.getenv("ENABLE_V1_KVCACHE_SCHEDULER", "0")):
                     self.max_num_batched_tokens = self.max_model_len
                 else:
-                    self.max_num_batched_tokens = 8192 # if set to max_model_len, it's easy to be OOM
+                    self.max_num_batched_tokens = 8192  # if set to max_model_len, it's easy to be OOM
 
         all_dict = asdict(self)
         all_dict["model_cfg"] = model_cfg
@@ -927,6 +947,7 @@ class EngineArgs:
             mm_processor_kwargs=self.mm_processor_kwargs,
             # enable_mm=self.enable_mm,
             reasoning_parser=self.reasoning_parser,
+            tool_parser=self.tool_call_parser,
             splitwise_role=self.splitwise_role,
             innode_prefill_ports=self.innode_prefill_ports,
             max_num_partial_prefills=self.max_num_partial_prefills,

fastdeploy/engine/config.py

@@ -86,6 +86,7 @@ class Config:
         max_long_partial_prefills: int = 1,
         long_prefill_token_threshold: int = 0,
         reasoning_parser: str = None,
+        tool_parser: str = None,
         guided_decoding_backend: Optional[str] = None,
         disable_any_whitespace: bool = False,
         enable_logprob: bool = False,
@@ -166,6 +167,7 @@ class Config:
         self.max_long_partial_prefills = max_long_partial_prefills
         self.long_prefill_token_threshold = long_prefill_token_threshold
         self.reasoning_parser = reasoning_parser
+        self.tool_parser = tool_parser
         self.graph_optimization_config = graph_optimization_config
         self.early_stop_config = early_stop_config
         self.guided_decoding_backend = guided_decoding_backend
@@ -245,10 +247,10 @@ class Config:
             if self.cache_config.enable_chunked_prefill:
                 self.max_num_batched_tokens = 2048
             else:
-                if not int(os.getenv('ENABLE_V1_KVCACHE_SCHEDULER', '0')):
+                if not int(os.getenv("ENABLE_V1_KVCACHE_SCHEDULER", "0")):
                     self.max_num_batched_tokens = self.max_model_len
                 else:
-                    self.max_num_batched_tokens = 8192 # if set to max_model_len, it's easy to be OOM
+                    self.max_num_batched_tokens = 8192  # if set to max_model_len, it's easy to be OOM
 
         if self.long_prefill_token_threshold == 0:
             self.long_prefill_token_threshold = int(self.max_model_len * 0.04)
@@ -296,7 +298,7 @@ class Config:
         )
 
         if not self.cache_config.enable_chunked_prefill:
-            if not int(os.getenv('ENABLE_V1_KVCACHE_SCHEDULER', '0')):
+            if not int(os.getenv("ENABLE_V1_KVCACHE_SCHEDULER", "0")):
                 assert self.max_num_batched_tokens >= self.max_model_len, (
                     f"max_num_batched_tokens: {self.max_num_batched_tokens} "
                     f"should be larger than or equal to max_model_len: {self.max_model_len}"

fastdeploy/engine/engine.py

@@ -106,6 +106,7 @@ class LLMEngine:
             cfg.limit_mm_per_prompt,
             cfg.mm_processor_kwargs,
             cfg.enable_mm,
+            cfg.tool_parser,
         )
 
         self.start_queue_service()

fastdeploy/engine/request.py

@@ -24,6 +24,7 @@ from typing import Any, Dict, Optional, Union
 import numpy as np
 
 from fastdeploy.engine.sampling_params import SamplingParams
+from fastdeploy.entrypoints.openai.protocol import ToolCall
 from fastdeploy.utils import data_processor_logger
 from fastdeploy.worker.output import LogprobsLists, SampleLogprobs
 
@@ -249,6 +250,7 @@ class CompletionOutput:
     draft_token_ids: list[int] = None
     text: Optional[str] = None
     reasoning_content: Optional[str] = None
+    tool_calls: Optional[ToolCall] = None
 
     def to_dict(self):
         """