hf_text-generation-inference/router/src/lib.rs

/// Text Generation Inference Webserver
mod infer;
mod queue;
pub mod server;
mod validation;

use infer::Infer;
use queue::{Entry, Queue};
use serde::{Deserialize, Serialize};
use utoipa::ToSchema;
use validation::Validation;

#[derive(Clone, Debug, Deserialize, ToSchema)]
pub(crate) struct GenerateParameters {
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        nullable = true,
        default = "null",
        example = 0.5
    )]
    pub temperature: Option<f32>,
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        nullable = true,
        default = "null",
        example = 1.03
    )]
    pub repetition_penalty: Option<f32>,
    #[serde(default)]
    #[schema(exclusive_minimum = 0, nullable = true, default = "null", example = 10)]
    pub top_k: Option<i32>,
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        maximum = 1.0,
        nullable = true,
        default = "null",
        example = 0.95
    )]
    pub top_p: Option<f32>,
    #[serde(default)]
    #[schema(
        exclusive_minimum = 0.0,
        maximum = 1.0,
        nullable = true,
        default = "null",
        example = 0.95
    )]
    pub typical_p: Option<f32>,
    #[serde(default)]
    #[schema(default = "false", example = true)]
    pub do_sample: bool,
    #[serde(default = "default_max_new_tokens")]
    #[schema(exclusive_minimum = 0, exclusive_maximum = 512, default = "20")]
    pub max_new_tokens: u32,
    #[serde(default)]
    #[schema(default = "None", example = false)]
    pub return_full_text: Option<bool>,
    #[serde(default)]
    #[schema(inline, max_items = 4, example = json ! (["photographer"]))]
    pub stop: Vec<String>,
    #[serde(default)]
    #[schema(default = "false", example = true)]
    pub watermark: bool,
    #[serde(default)]
    #[schema(default = "true")]
    pub details: bool,
    #[serde(default)]
    pub seed: Option<u64>,
}

fn default_max_new_tokens() -> u32 {
    20
}

fn default_parameters() -> GenerateParameters {
    GenerateParameters {
        temperature: None,
        repetition_penalty: None,
        top_k: None,
        top_p: None,
        typical_p: None,
        do_sample: false,
        max_new_tokens: default_max_new_tokens(),
        return_full_text: None,
        stop: Vec::new(),
        watermark: false,
        details: false,
        seed: None,
    }
}

#[derive(Clone, Debug, Deserialize, ToSchema)]
pub(crate) struct GenerateRequest {
    #[schema(example = "My name is Olivier and I")]
    pub inputs: String,
    #[serde(default = "default_parameters")]
    pub parameters: GenerateParameters,
}

#[derive(Clone, Debug, Deserialize, ToSchema)]
pub(crate) struct CompatGenerateRequest {
    #[schema(example = "My name is Olivier and I")]
    pub inputs: String,
    #[serde(default = "default_parameters")]
    pub parameters: GenerateParameters,
    #[serde(default)]
    #[allow(dead_code)]
    pub stream: bool,
}

impl From<CompatGenerateRequest> for GenerateRequest {
    fn from(req: CompatGenerateRequest) -> Self {
        Self {
            inputs: req.inputs,
            parameters: req.parameters,
        }
    }
}

#[derive(Debug, Serialize, ToSchema)]
pub struct PrefillToken {
    #[schema(example = 0)]
    id: u32,
    #[schema(example = "test")]
    text: String,
    #[schema(nullable = true, example = - 0.34)]
    logprob: f32,
}

#[derive(Debug, Serialize, ToSchema)]
pub struct Token {
    #[schema(example = 0)]
    id: u32,
    #[schema(example = "test")]
    text: String,
    #[schema(nullable = true, example = - 0.34)]
    logprob: f32,
    #[schema(example = "false")]
    special: bool,
}

#[derive(Serialize, ToSchema)]
#[serde(rename_all(serialize = "snake_case"))]
pub(crate) enum FinishReason {
    #[schema(rename = "length")]
    Length,
    #[serde(rename = "eos_token")]
    #[schema(rename = "eos_token")]
    EndOfSequenceToken,
    #[schema(rename = "stop_sequence")]
    StopSequence,
}

#[derive(Serialize, ToSchema)]
pub(crate) struct Details {
    #[schema(example = "length")]
    pub finish_reason: FinishReason,
    #[schema(example = 1)]
    pub generated_tokens: u32,
    #[schema(example = 42)]
    pub seed: Option<u64>,
    pub prefill: Vec<PrefillToken>,
    pub tokens: Vec<Token>,
}

#[derive(Serialize, ToSchema)]
pub(crate) struct GenerateResponse {
    #[schema(example = "test")]
    pub generated_text: String,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub details: Option<Details>,
}

#[derive(Serialize, ToSchema)]
pub(crate) struct StreamDetails {
    #[schema(example = "length")]
    pub finish_reason: FinishReason,
    #[schema(example = 1)]
    pub generated_tokens: u32,
    #[schema(example = 42)]
    pub seed: Option<u64>,
}

#[derive(Serialize, ToSchema)]
pub(crate) struct StreamResponse {
    pub token: Token,
    #[schema(nullable = true, default = "null", example = "test")]
    pub generated_text: Option<String>,
    #[schema(nullable = true, default = "null")]
    pub details: Option<StreamDetails>,
}

#[derive(Serialize, ToSchema)]
pub(crate) struct ErrorResponse {
    pub error: String,
    pub error_type: String,
}