File size: 10,456 Bytes
bcdf9fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright 2023-2024 SGLang Team
# Copyright 2025 ModelBest Inc. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from enum import Enum
from typing import Any, Dict, List, Literal, Optional

import torch
from pydantic import BaseModel
from transformers import PreTrainedTokenizer

from verl.tools.schemas import OpenAIFunctionToolCall, OpenAIFunctionToolSchema
from verl.utils.model import compute_position_id_with_mask


class FinishReasonTypeEnum(str, Enum):
    """The enum for finish reason type."""

    LENGTH = "length"
    STOP = "stop"
    TOOL_CALL = "tool_calls"

    @classmethod
    def from_str(cls, value: str) -> "FinishReasonTypeEnum":
        if value == "stop":
            return cls.STOP
        elif value == "length":
            return cls.LENGTH
        elif value == "tool_calls":
            return cls.TOOL_CALL
        else:
            raise ValueError(f"Unsupported finish reason type: {value}")


class Message(BaseModel):
    role: str
    content: str
    tool_calls: Optional[List[OpenAIFunctionToolCall]] = None


class AsyncRolloutRequestStateEnum(str, Enum):
    """The enum for async rollout request state."""

    PENDING = "pending"
    RUNNING = "running"
    COMPLETED = "completed"
    FAILED = "failed"
    TOOL_CALLING = "tool_calling"


class AsyncRolloutRequest(BaseModel):
    """The data model for async rollout."""

    batch_data_id: int = 0
    rollout_offset: int = 0
    request_id: str
    state: AsyncRolloutRequestStateEnum
    messages: List[Message]
    tools: Optional[List[OpenAIFunctionToolSchema]] = None
    tools_kwargs: Dict[str, Any] = {}
    input_ids: List[int]
    prompt_ids: List[int]
    response_ids: List[int]
    attention_mask: List[int]
    prompt_attention_mask: List[int]
    response_attention_mask: List[int]
    position_ids: List[int]
    prompt_position_ids: List[int]
    response_position_ids: List[int]
    loss_mask: List[int]
    prompt_loss_mask: List[int]
    response_loss_mask: List[int]
    reward_scores: Dict[str, float]
    max_response_len: int = 8192
    max_model_len: int = 32768

    format_config: dict = {
        "chatml": {
            "assistant_prefix_msg": "\n<|im_start|>assistant\n",
            "assistant_suffix_msg": "<|im_end|>",
            "tool_prefix_msg": "\n<|im_start|>tool\n",
            "tool_suffix_msg": "<|im_end|>",
        }
    }

    def get_generation_prompt(self, tokenizer: PreTrainedTokenizer) -> str:
        return tokenizer.apply_chat_template(  # type: ignore
            conversation=[msg.model_dump() for msg in self.messages],
            tools=[tool.model_dump() for tool in self.tools] if self.tools else None,
            add_generation_prompt=True,
            tokenize=False,
        )

    def add_assistant_message(

        self,

        tokenizer: PreTrainedTokenizer,

        content: str,

        tool_calls: Optional[List[OpenAIFunctionToolCall]] = None,

        format: Literal["chatml"] = "chatml",

        already_over_long: bool = False,

    ) -> None:
        """Currently, we only support chatml format."""
        msg = Message(role="assistant", content=content, tool_calls=tool_calls)
        self.messages.append(msg)
        if tool_calls is not None:
            content_with_tool_calls: str = tokenizer.apply_chat_template(  # type: ignore
                conversation=[msg.model_dump()], add_generation_prompt=False, tokenize=False
            )
        else:
            content_with_tool_calls = content
        # TODO: support other formats
        if format in self.format_config:
            prefix_msg = self.format_config[format]["assistant_prefix_msg"]
            prefix_token_ids = tokenizer.encode(prefix_msg, add_special_tokens=False)
            suffix_msg = self.format_config[format]["assistant_suffix_msg"]
            suffix_token_ids = tokenizer.encode(suffix_msg, add_special_tokens=False)
            if tool_calls is not None:
                content = content_with_tool_calls.split(f"{prefix_msg}")[-1].split(f"{suffix_msg}")[0]
            content_token_ids = tokenizer.encode(content, add_special_tokens=False)
            if self.input_ids[-len(prefix_token_ids) :] == prefix_token_ids:
                append_token_ids = content_token_ids
                _loss_mask = [1] * len(content_token_ids)
            elif self.input_ids[-len(suffix_token_ids) :] == suffix_token_ids:
                append_token_ids = prefix_token_ids + content_token_ids
                _loss_mask = [0] * len(prefix_token_ids) + [1] * len(content_token_ids)
            else:
                max_len = max(len(prefix_token_ids), len(suffix_token_ids))
                raise ValueError(
                    f"""Unsupported end of message format: 

                    {tokenizer.decode(self.input_ids[-max_len:])},

                    {tokenizer.decode(self.input_ids)=}, {self.messages=}"""
                )
            if not already_over_long:
                append_token_ids += suffix_token_ids
                _loss_mask += [1] * len(suffix_token_ids)
            self.input_ids += append_token_ids
            _attention_mask = [1] * len(append_token_ids)
            self.attention_mask += _attention_mask
            _delta_position_ids = compute_position_id_with_mask(torch.tensor(_attention_mask)).tolist()
            last_position_id = self.position_ids[-1]
            _position_ids = [pos_id + last_position_id for pos_id in _delta_position_ids]
            self.loss_mask += _loss_mask
            self.position_ids += _position_ids
        else:
            raise ValueError(f"Unsupported format: {format}")
        assert len(self.input_ids) == len(self.attention_mask) == len(self.position_ids) == len(self.loss_mask), f"""Request {self.request_id} has different length of {len(self.input_ids)=}, 

            {len(self.attention_mask)=}, {len(self.position_ids)=}, {len(self.loss_mask)=}"""

    def add_tool_response_message(self, tokenizer: PreTrainedTokenizer, content: str, format: Literal["chatml"] = "chatml") -> None:
        """Currently, we only support chatml format."""
        msg = Message(role="tool", content=content)
        self.messages.append(msg)
        # TODO: support other formats
        if format in self.format_config:
            prefix_msg = self.format_config[format]["tool_prefix_msg"]
            prefix_token_ids = tokenizer.encode(prefix_msg, add_special_tokens=False)
            suffix_msg = self.format_config[format]["tool_suffix_msg"]
            suffix_token_ids = tokenizer.encode(suffix_msg, add_special_tokens=False)
            content_token_ids = tokenizer.encode(content, add_special_tokens=False)
            if self.input_ids[-len(prefix_token_ids) :] == prefix_token_ids:
                append_token_ids = content_token_ids + suffix_token_ids
            elif self.input_ids[-len(suffix_token_ids) :] == suffix_token_ids:
                append_token_ids = prefix_token_ids + content_token_ids + suffix_token_ids
            else:
                raise ValueError(f"Unsupported end of message format: {tokenizer.decode(self.input_ids[-len(prefix_token_ids) :])}")
            self.input_ids += append_token_ids
            _attention_mask = [1] * len(append_token_ids)
            self.attention_mask += _attention_mask
            _delta_position_ids = compute_position_id_with_mask(torch.tensor(_attention_mask)).tolist()
            last_position_id = self.position_ids[-1]
            _position_ids = [pos_id + last_position_id for pos_id in _delta_position_ids]
            self.loss_mask += [0] * len(append_token_ids)
            self.position_ids += _position_ids
        else:
            raise ValueError(f"Unsupported format: {format}")
        assert len(self.input_ids) == len(self.attention_mask) == len(self.position_ids) == len(self.loss_mask), f"""Request {self.request_id} has different length of {len(self.input_ids)=}, 

            {len(self.attention_mask)=}, {len(self.position_ids)=}, {len(self.loss_mask)=}"""

    def finalize(

        self,

        tokenizer: PreTrainedTokenizer,

        reward_scores: Dict[str, float],

        finish_reason_type: FinishReasonTypeEnum = FinishReasonTypeEnum.STOP,

    ) -> None:
        self.state = AsyncRolloutRequestStateEnum.COMPLETED
        self.reward_scores = reward_scores
        self.response_ids = self.input_ids[len(self.prompt_ids) :]
        if finish_reason_type == FinishReasonTypeEnum.STOP:
            pass
        elif finish_reason_type == FinishReasonTypeEnum.LENGTH:
            pass
        else:
            raise ValueError(f"Unsupported finalize finish reason type: {finish_reason_type}")
        self.truncate_output_ids(tokenizer)
        assert len(self.input_ids) == len(self.attention_mask) == len(self.position_ids) == len(self.loss_mask), f"""Request {self.request_id} has different length of {len(self.input_ids)=}, 

            {len(self.attention_mask)=}, {len(self.position_ids)=}, {len(self.loss_mask)=}"""

    def truncate_output_ids(self, tokenizer: PreTrainedTokenizer) -> None:
        self.input_ids = self.input_ids[: self.max_model_len]
        self.attention_mask = self.attention_mask[: self.max_model_len]
        self.position_ids = self.position_ids[: self.max_model_len]
        self.loss_mask = self.loss_mask[: self.max_model_len]
        self.response_ids = self.input_ids[len(self.prompt_ids) :][: self.max_response_len]
        self.response_attention_mask = self.attention_mask[len(self.prompt_attention_mask) :][: self.max_response_len]
        self.response_position_ids = self.position_ids[len(self.prompt_position_ids) :][: self.max_response_len]
        self.response_loss_mask = self.loss_mask[len(self.prompt_loss_mask) :][: self.max_response_len]