nvext.rs 4.22 KB
Newer Older
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
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

use derive_builder::Builder;
use serde::{Deserialize, Serialize};
use validator::{Validate, ValidationError};

pub trait NvExtProvider {
    fn nvext(&self) -> Option<&NvExt>;
    fn raw_prompt(&self) -> Option<String>;
}

/// NVIDIA LLM extensions to the OpenAI API
#[derive(Serialize, Deserialize, Builder, Validate, Debug, Clone)]
#[validate(schema(function = "validate_nv_ext"))]
pub struct NvExt {
    /// If true, sampling will be forced to be greedy.
    /// The backend is responsible for selecting the correct backend-specific options to
    /// implement this.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    #[builder(default, setter(strip_option))]
    pub greed_sampling: Option<bool>,

    /// If true, the preproessor will try to bypass the prompt template and pass the prompt directly to
    /// to the tokenizer.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    #[builder(default, setter(strip_option))]
    pub use_raw_prompt: Option<bool>,

    /// Annotations
    /// User requests triggers which result in the request issue back out-of-band information in the SSE
    /// stream using the `event:` field.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    #[builder(default, setter(strip_option))]
    pub annotations: Option<Vec<String>>,
36

37
38
39
40
41
42
    /// Targeted backend instance ID for the request
    /// If set, the request will be routed to backend instance with the given ID.
    /// If not set, the request will be routed to the best matching instance.
    #[builder(default, setter(strip_option))]
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub backend_instance_id: Option<i64>,
43
44
45
46
47
48
49

    /// Pre-tokenized data to use instead of tokenizing the prompt
    /// If provided along with backend_instance_id, these tokens will be used directly
    /// and tokenization will be skipped.
    #[builder(default, setter(strip_option))]
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub token_data: Option<Vec<u32>>,
50

51
52
53
54
55
    /// Maximum number of thinking tokens allowed
    /// NOTE: Currently passed through to backends as a no-op for future implementation
    #[serde(default, skip_serializing_if = "Option::is_none")]
    #[builder(default, setter(strip_option))]
    pub max_thinking_tokens: Option<u32>,
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
}

impl Default for NvExt {
    fn default() -> Self {
        NvExt::builder().build().unwrap()
    }
}

impl NvExt {
    pub fn builder() -> NvExtBuilder {
        NvExtBuilder::default()
    }
}

fn validate_nv_ext(_nv_ext: &NvExt) -> Result<(), ValidationError> {
    Ok(())
}

impl NvExtBuilder {
    pub fn add_annotation(&mut self, annotation: impl Into<String>) -> &mut Self {
        self.annotations
            .get_or_insert_with(|| Some(vec![]))
            .as_mut()
            .expect("stop should always be Some(Vec)")
            .push(annotation.into());
        self
    }
}

#[cfg(test)]
mod tests {
    use validator::Validate;

    use super::*;

    // Test default builder configuration
    #[test]
    fn test_nv_ext_builder_default() {
        let nv_ext = NvExt::builder().build().unwrap();
        assert_eq!(nv_ext.greed_sampling, None);
96
97
98
99
        assert_eq!(nv_ext.use_raw_prompt, None);
        assert_eq!(nv_ext.annotations, None);
        assert_eq!(nv_ext.backend_instance_id, None);
        assert_eq!(nv_ext.token_data, None);
100
        assert_eq!(nv_ext.max_thinking_tokens, None);
101
102
103
104
105
106
107
    }

    // Test valid builder configurations
    #[test]
    fn test_nv_ext_builder_custom() {
        let nv_ext = NvExt::builder()
            .greed_sampling(true)
108
109
110
            .use_raw_prompt(true)
            .backend_instance_id(42)
            .token_data(vec![1, 2, 3, 4])
111
            .max_thinking_tokens(1024)
112
113
114
115
            .build()
            .unwrap();

        assert_eq!(nv_ext.greed_sampling, Some(true));
116
117
118
        assert_eq!(nv_ext.use_raw_prompt, Some(true));
        assert_eq!(nv_ext.backend_instance_id, Some(42));
        assert_eq!(nv_ext.token_data, Some(vec![1, 2, 3, 4]));
119
        assert_eq!(nv_ext.max_thinking_tokens, Some(1024));
120
121
122
123
        // Validate the built struct
        assert!(nv_ext.validate().is_ok());
    }
}