nuts_rs/

sampler.rs

//! High-level sampler entry points: `Settings` presets, the parallel `Sampler`,
//! and `sample_sequentially` for running one or many chains.

use anyhow::Result;
use nuts_storable::{HasDims, Storable, Value};
use rand::{Rng, SeedableRng, rngs::ChaCha8Rng};
use serde::{Deserialize, Serialize, de::DeserializeOwned};
use std::{collections::HashMap, fmt::Debug, time::Duration};

#[cfg(feature = "parallel")]
use anyhow::{Context, bail};
#[cfg(feature = "parallel")]
use itertools::Itertools;
#[cfg(feature = "parallel")]
use std::ops::Deref;

#[cfg(feature = "parallel")]
use rayon::{ScopeFifo, ThreadPoolBuilder};
#[cfg(feature = "parallel")]
use std::{
    sync::{
        Arc, Mutex,
        mpsc::{
            Receiver, RecvTimeoutError, Sender, SyncSender, TryRecvError, channel, sync_channel,
        },
    },
    thread::{JoinHandle, spawn},
    time::Instant,
};

use crate::{
    DiagAdaptExpSettings, Math, StepSizeAdaptMethod,
    adapt_strategy::{EuclideanAdaptOptions, GlobalStrategy, GlobalStrategyStatsOptions},
    chain::{AdaptStrategy, Chain, NutsChain, StatOptions},
    dynamics::{KineticEnergyKind, TransformedHamiltonian, TransformedPointStatsOptions},
    external_adapt_strategy::{ExternalTransformAdaptation, FlowSettings},
    mclmc::MclmcTrajectoryKind,
    nuts::NutsOptions,
    sampler_stats::{SamplerStats, StatsDims},
    transform::{
        DiagAdaptStrategy, DiagMassMatrix, ExternalTransformation, LowRankMassMatrix,
        LowRankMassMatrixStrategy, LowRankSettings,
    },
};

#[cfg(feature = "parallel")]
use crate::{
    model::Model,
    storage::{ChainStorage, StorageConfig, TraceStorage},
};

/// All sampler configurations implement this trait
pub trait Settings:
    private::Sealed + Clone + Copy + Default + Sync + Send + Serialize + DeserializeOwned + 'static
{
    type Chain<M: Math>: Chain<M>;

    fn new_chain<M: Math, R: Rng + ?Sized>(
        &self,
        chain: u64,
        math: M,
        rng: &mut R,
    ) -> Self::Chain<M>;

    fn hint_num_tune(&self) -> usize;
    fn hint_num_draws(&self) -> usize;
    fn num_chains(&self) -> usize;
    fn seed(&self) -> u64;
    fn stats_options<M: Math>(&self) -> <Self::Chain<M> as SamplerStats<M>>::StatsOptions;
    fn sampler_name(&self) -> &'static str;
    fn adaptation_name(&self) -> &'static str;

    fn stat_names<M: Math>(&self, math: &M) -> Vec<String> {
        let dims = StatsDims::from(math);
        <<Self::Chain<M> as SamplerStats<M>>::Stats as Storable<_>>::names(&dims)
            .into_iter()
            .map(String::from)
            .collect()
    }

    fn data_names<M: Math>(&self, math: &M) -> Vec<String> {
        <M::ExpandedVector as Storable<_>>::names(math)
            .into_iter()
            .map(String::from)
            .collect()
    }

    fn stat_types<M: Math>(&self, math: &M) -> Vec<(String, nuts_storable::ItemType)> {
        self.stat_names(math)
            .into_iter()
            .map(|name| (name.clone(), self.stat_type::<M>(math, &name)))
            .collect()
    }

    fn stat_type<M: Math>(&self, math: &M, name: &str) -> nuts_storable::ItemType {
        let dims = StatsDims::from(math);
        <<Self::Chain<M> as SamplerStats<M>>::Stats as Storable<_>>::item_type(&dims, name)
    }

    fn data_types<M: Math>(&self, math: &M) -> Vec<(String, nuts_storable::ItemType)> {
        self.data_names(math)
            .into_iter()
            .map(|name| (name.clone(), self.data_type(math, &name)))
            .collect()
    }
    fn data_type<M: Math>(&self, math: &M, name: &str) -> nuts_storable::ItemType {
        <M::ExpandedVector as Storable<_>>::item_type(math, name)
    }

    fn stat_dims_all<M: Math>(&self, math: &M) -> Vec<(String, Vec<String>)> {
        self.stat_names(math)
            .into_iter()
            .map(|name| (name.clone(), self.stat_dims::<M>(math, &name)))
            .collect()
    }

    fn stat_dims<M: Math>(&self, math: &M, name: &str) -> Vec<String> {
        let dims = StatsDims::from(math);
        <<Self::Chain<M> as SamplerStats<M>>::Stats as Storable<_>>::dims(&dims, name)
            .into_iter()
            .map(String::from)
            .collect()
    }

    fn stat_dim_sizes<M: Math>(&self, math: &M) -> HashMap<String, u64> {
        let dims = StatsDims::from(math);
        dims.dim_sizes()
    }

    fn data_dims_all<M: Math>(&self, math: &M) -> Vec<(String, Vec<String>)> {
        self.data_names(math)
            .into_iter()
            .map(|name| (name.clone(), self.data_dims(math, &name)))
            .collect()
    }

    fn data_dims<M: Math>(&self, math: &M, name: &str) -> Vec<String> {
        <M::ExpandedVector as Storable<_>>::dims(math, name)
            .into_iter()
            .map(String::from)
            .collect()
    }

    fn stat_coords<M: Math>(&self, math: &M) -> HashMap<String, Value> {
        let dims = StatsDims::from(math);
        dims.coords()
    }

    fn stat_event_dims<M: Math>(&self, math: &M) -> Vec<(String, Option<String>)> {
        let dims = StatsDims::from(math);
        self.stat_names(math)
            .into_iter()
            .map(|name| {
                let event_dim =
                    <<Self::Chain<M> as SamplerStats<M>>::Stats as Storable<_>>::event_dim(
                        &dims, &name,
                    )
                    .map(String::from);
                (name, event_dim)
            })
            .collect()
    }
}

#[derive(Debug, Clone)]
#[non_exhaustive]
pub struct Progress {
    pub draw: u64,
    pub chain: u64,
    pub diverging: bool,
    pub tuning: bool,
    pub step_size: f64,
    pub num_steps: u64,
}

mod private {
    use super::{
        DiagMclmcSettings, DiagNutsSettings, FlowMclmcSettings, FlowNutsSettings,
        LowRankMclmcSettings, LowRankNutsSettings,
    };

    pub trait Sealed {}

    impl Sealed for DiagNutsSettings {}

    impl Sealed for LowRankNutsSettings {}

    impl Sealed for FlowNutsSettings {}

    impl Sealed for DiagMclmcSettings {}

    impl Sealed for LowRankMclmcSettings {}

    impl Sealed for FlowMclmcSettings {}
}

/// Settings for the NUTS sampler
#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
pub struct NutsSettings<A: Debug + Copy + Default + Serialize> {
    /// The number of tuning steps, where we fit the step size and geometry.
    pub num_tune: u64,
    /// The number of draws after tuning
    pub num_draws: u64,
    /// The maximum tree depth during sampling. The number of leapfrog steps
    /// is smaller than 2 ^ maxdepth.
    pub maxdepth: u64,
    /// The minimum tree depth during sampling. The number of leapfrog steps
    /// is larger than 2 ^ mindepth.
    pub mindepth: u64,
    /// Store the gradient in the SampleStats
    pub store_gradient: bool,
    /// Store each unconstrained parameter vector in the sampler stats
    pub store_unconstrained: bool,
    /// Store the transformed gradient and value in the sampler stats
    pub store_transformed: bool,
    /// If the energy error is larger than this threshold we treat the leapfrog
    /// step as a divergence.
    pub max_energy_error: f64,
    /// Store detailed information about each divergence in the sampler stats
    pub store_divergences: bool,
    /// Settings for geometry adaptation.
    pub adapt_options: A,
    pub check_turning: bool,
    pub target_integration_time: Option<f64>,
    /// Selects the kinetic-energy form and the corresponding integrator.
    ///
    /// - [`KineticEnergyKind::Euclidean`]: standard leapfrog (default for most settings).
    /// - [`KineticEnergyKind::ExactNormal`]: geodesic leapfrog exact for a standard-normal
    ///   potential.
    /// - [`KineticEnergyKind::Microcanonical`]: isokinetic ESH-dynamics leapfrog (microcanonical
    ///   HMC); momentum is constrained to the unit sphere.
    pub trajectory_kind: KineticEnergyKind,
    pub num_chains: usize,
    pub seed: u64,
    /// Number of extra doublings to perform after reaching maxdepth. This can
    /// be used to increase the effective sample size at the cost of more
    /// expensive sampling.
    pub extra_doublings: u64,
}

pub type DiagNutsSettings = NutsSettings<EuclideanAdaptOptions<DiagAdaptExpSettings>>;
/// Backwards-compatible alias for [`DiagNutsSettings`].
#[deprecated(since = "0.0.0", note = "Use DiagNutsSettings instead")]
pub type DiagGradNutsSettings = DiagNutsSettings;
pub type LowRankNutsSettings = NutsSettings<EuclideanAdaptOptions<LowRankSettings>>;
pub type FlowNutsSettings = NutsSettings<FlowSettings>;
/// Backwards-compatible alias for [`FlowNutsSettings`].
#[deprecated(since = "0.0.0", note = "Use FlowNutsSettings instead")]
pub type TransformedNutsSettings = FlowNutsSettings;

/// Settings for the unadjusted Microcanonical Langevin Monte Carlo (MCLMC) sampler.
///
/// > ⚠️ **Experimental — use with caution**: The MCLMC sampler and all of its
/// > variants are highly experimental. They have not been thoroughly validated
/// > and may **not return correct posteriors**. The API, defaults, and
/// > adaptation behaviour are all subject to breaking changes at any time.
/// > Do not use these samplers in production or for results you rely on.
///
/// Step size `ε` and momentum decoherence length `L` are **constants** — no
/// adaptation of those is performed yet. The geometry is adapted during
/// warmup using the sampler-specific adaptation strategy, while the step size
/// remains fixed.
///
/// Use the type aliases [`DiagMclmcSettings`], [`LowRankMclmcSettings`], and
/// [`FlowMclmcSettings`] for concrete configurations.
#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
pub struct MclmcSettings<A: Debug + Copy + Default + Serialize> {
    /// Step size ε for the ESH leapfrog integrator.
    pub step_size: f64,
    /// Momentum decoherence length L (controls partial momentum refresh rate).
    /// Set to `f64::INFINITY` to disable momentum refresh entirely.
    pub momentum_decoherence_length: f64,
    /// Number of warmup draws.
    pub num_tune: u64,
    /// Number of sampling draws after warmup.
    pub num_draws: u64,
    /// Number of parallel chains.
    pub num_chains: usize,
    /// RNG seed.
    pub seed: u64,
    /// Maximum energy error before a step is flagged as a divergence.
    pub max_energy_error: f64,
    /// Store each unconstrained parameter vector in the sampler stats.
    pub store_unconstrained: bool,
    /// Store the gradient in the sampler stats.
    pub store_gradient: bool,
    /// Store the transformed gradient and value in the sampler stats
    pub store_transformed: bool,
    /// Store detailed information about each divergence in the sampler stats
    pub store_divergences: bool,
    /// Geometry adaptation options (step-size fields are ignored for Euclidean settings).
    pub adapt_options: A,
    /// Number of leapfrog steps per draw as a fraction of `L / ε`.
    ///
    /// The number of leapfrog steps between collector calls is:
    /// `round(subsample_frequency * L / ε).max(1)`
    ///
    /// - `1.0` (default) — one sample per full trajectory (at the final step).
    /// - `0.0` — every leapfrog step.
    /// - Values in between space samples as a fraction of the decoherence
    ///   length, so the interval scales naturally when `L` or `ε` changes.
    pub subsample_frequency: f64,
    /// When `true`, use the tree-structured step size retry on divergence:
    /// halve the step size factor and try 2 steps before doubling back.
    /// `log_weight` will include `log(step_size)` to correct for the varying
    /// sampling density. When `false`, divergences are recorded immediately
    /// without any retry and `log_weight = -energy_change`.
    pub dynamic_step_size: bool,
    /// Selects which leapfrog integrator and partial-momentum-refresh style
    /// to use.  See [`MclmcTrajectoryKind`] for the available options.
    /// Default: [`MclmcTrajectoryKind::Microcanonical`] (original MCLMC).
    pub trajectory_kind: MclmcTrajectoryKind,
    /// Fraction of `num_tune` draws at which the trajectory is switched from
    /// Euclidean to Microcanonical when
    /// `trajectory_kind == MclmcTrajectoryKind::EuclideanEarlyThenMicrocanonical`.
    /// Ignored for other trajectory kinds.  Default: `0.3`.
    pub trajectory_switch_fraction: f64,
}

/// MCLMC settings with a diagonal mass matrix adaptation.
///
/// > ⚠️ **Experimental — use with caution**: Highly experimental. Correctness
/// > of the returned posteriors has not been verified. May change at any time.
pub type DiagMclmcSettings = MclmcSettings<EuclideanAdaptOptions<DiagAdaptExpSettings>>;
/// MCLMC settings with a low-rank mass matrix adaptation.
///
/// > ⚠️ **Experimental — use with caution**: Highly experimental. Correctness
/// > of the returned posteriors has not been verified. May change at any time.
pub type LowRankMclmcSettings = MclmcSettings<EuclideanAdaptOptions<LowRankSettings>>;
/// MCLMC settings with a learned flow transformation.
///
/// > ⚠️ **Experimental — use with caution**: Highly experimental. Correctness
/// > of the returned posteriors has not been verified. May change at any time.
pub type FlowMclmcSettings = MclmcSettings<FlowSettings>;
/// Backwards-compatible alias for [`FlowMclmcSettings`].
#[deprecated(since = "0.0.0", note = "Use FlowMclmcSettings instead")]
pub type TransformedMclmcSettings = FlowMclmcSettings;

fn usize_hint(value: u64, field: &str) -> usize {
    value
        .try_into()
        .unwrap_or_else(|_| panic!("{field} must be smaller than usize::MAX"))
}

fn default_mclmc_settings<A: Debug + Copy + Default + Serialize>(
    adapt_options: A,
    num_tune: u64,
    num_chains: usize,
    max_energy_error: f64,
) -> MclmcSettings<A> {
    MclmcSettings {
        step_size: 0.5,
        momentum_decoherence_length: 3.0,
        num_tune,
        num_draws: 1000,
        num_chains,
        seed: 0,
        max_energy_error,
        store_unconstrained: false,
        store_gradient: false,
        store_divergences: false,
        store_transformed: false,
        adapt_options,
        subsample_frequency: 1.0,
        dynamic_step_size: true,
        trajectory_kind: MclmcTrajectoryKind::EuclideanEarlyThenMicrocanonical,
        trajectory_switch_fraction: 0.3,
    }
}

impl Default for DiagMclmcSettings {
    fn default() -> Self {
        let mut adapt_options = EuclideanAdaptOptions::default();
        adapt_options.step_size_settings.adapt_options.method = StepSizeAdaptMethod::Fixed(0.5);
        default_mclmc_settings(adapt_options, 400, 6, 1000.0)
    }
}

impl Default for LowRankMclmcSettings {
    fn default() -> Self {
        let mut adapt_options = EuclideanAdaptOptions::default();
        adapt_options.early_mass_matrix_switch_freq = 20;
        adapt_options.step_size_settings.adapt_options.method = StepSizeAdaptMethod::Fixed(0.5);
        default_mclmc_settings(adapt_options, 800, 6, 1000.0)
    }
}

impl Default for FlowMclmcSettings {
    fn default() -> Self {
        default_mclmc_settings(FlowSettings::default(), 1500, 1, 20.0)
    }
}

type DiagMclmcChain<M> = crate::mclmc::MclmcChain<
    M,
    ChaCha8Rng,
    GlobalStrategy<M, DiagAdaptStrategy<M>>,
    DiagMassMatrix<M>,
>;
type LowRankMclmcChain<M> = crate::mclmc::MclmcChain<
    M,
    ChaCha8Rng,
    GlobalStrategy<M, LowRankMassMatrixStrategy>,
    LowRankMassMatrix<M>,
>;

impl Settings for DiagMclmcSettings {
    type Chain<M: Math> = DiagMclmcChain<M>;

    fn new_chain<M: Math, R: Rng + ?Sized>(
        &self,
        chain: u64,
        mut math: M,
        rng: &mut R,
    ) -> Self::Chain<M> {
        use crate::dynamics::KineticEnergyKind;
        use crate::mclmc::MclmcChain;
        use crate::stepsize::StepSizeAdaptMethod;

        let num_tune = self.num_tune;
        let mut adapt_options = self.adapt_options;
        adapt_options.step_size_settings.adapt_options.method =
            StepSizeAdaptMethod::Fixed(self.step_size);
        let strategy = GlobalStrategy::<M, DiagAdaptStrategy<M>>::new(
            &mut math,
            adapt_options,
            num_tune,
            chain,
        );
        let mass_matrix = DiagMassMatrix::new(
            &mut math,
            self.adapt_options.mass_matrix_options.store_mass_matrix,
        );
        let initial_kind = match self.trajectory_kind {
            MclmcTrajectoryKind::Microcanonical => KineticEnergyKind::Microcanonical,
            MclmcTrajectoryKind::Euclidean
            | MclmcTrajectoryKind::EuclideanEarlyThenMicrocanonical => KineticEnergyKind::Euclidean,
        };
        let mut hamiltonian = TransformedHamiltonian::new(&mut math, mass_matrix, initial_kind);
        hamiltonian.set_momentum_decoherence_length(Some(self.momentum_decoherence_length));
        let switch_draw = (self.trajectory_switch_fraction * self.num_tune as f64) as u64;
        let rng = ChaCha8Rng::try_from_rng(rng).expect("Could not seed rng");
        let stats_options = self.stats_options::<M>();
        MclmcChain::new(
            math,
            hamiltonian,
            strategy,
            rng,
            chain,
            self.subsample_frequency,
            self.dynamic_step_size,
            self.trajectory_kind,
            switch_draw,
            self.max_energy_error,
            stats_options,
        )
    }

    fn hint_num_tune(&self) -> usize {
        usize_hint(self.num_tune, "num_tune")
    }

    fn hint_num_draws(&self) -> usize {
        usize_hint(self.num_draws, "num_draws")
    }

    fn num_chains(&self) -> usize {
        self.num_chains
    }

    fn seed(&self) -> u64 {
        self.seed
    }

    fn stats_options<M: Math>(&self) -> <Self::Chain<M> as SamplerStats<M>>::StatsOptions {
        StatOptions {
            adapt: GlobalStrategyStatsOptions {
                step_size: (),
                mass_matrix: (),
            },
            hamiltonian: -1,
            point: {
                let store_gradient = self.store_gradient;
                let store_unconstrained = self.store_unconstrained;
                let store_transformed = self.store_transformed;
                TransformedPointStatsOptions {
                    store_gradient,
                    store_unconstrained,
                    store_transformed,
                }
            },
            divergence: crate::dynamics::DivergenceStatsOptions {
                store_divergences: self.store_divergences,
            },
        }
    }

    fn sampler_name(&self) -> &'static str {
        "mclmc"
    }

    fn adaptation_name(&self) -> &'static str {
        "diagonal"
    }
}

fn default_nuts_settings<A: Debug + Copy + Default + Serialize>(
    adapt_options: A,
    num_tune: u64,
    num_chains: usize,
    max_energy_error: f64,
) -> NutsSettings<A> {
    NutsSettings {
        num_tune,
        num_draws: 1000,
        maxdepth: 10,
        mindepth: 0,
        max_energy_error,
        store_gradient: false,
        store_unconstrained: false,
        store_transformed: false,
        store_divergences: false,
        adapt_options,
        check_turning: true,
        seed: 0,
        num_chains,
        target_integration_time: None,
        trajectory_kind: KineticEnergyKind::Euclidean,
        extra_doublings: 0,
    }
}

impl Settings for LowRankMclmcSettings {
    type Chain<M: Math> = LowRankMclmcChain<M>;

    fn new_chain<M: Math, R: Rng + ?Sized>(
        &self,
        chain: u64,
        mut math: M,
        rng: &mut R,
    ) -> Self::Chain<M> {
        use crate::dynamics::KineticEnergyKind;
        use crate::mclmc::MclmcChain;
        use crate::stepsize::StepSizeAdaptMethod;

        let num_tune = self.num_tune;
        let mut adapt_options = self.adapt_options;
        adapt_options.step_size_settings.adapt_options.method =
            StepSizeAdaptMethod::Fixed(self.step_size);
        let strategy = GlobalStrategy::<M, LowRankMassMatrixStrategy>::new(
            &mut math,
            adapt_options,
            num_tune,
            chain,
        );
        let mass_matrix = LowRankMassMatrix::new(&mut math, self.adapt_options.mass_matrix_options);
        let initial_kind = match self.trajectory_kind {
            MclmcTrajectoryKind::Microcanonical => KineticEnergyKind::Microcanonical,
            MclmcTrajectoryKind::Euclidean
            | MclmcTrajectoryKind::EuclideanEarlyThenMicrocanonical => KineticEnergyKind::Euclidean,
        };
        let mut hamiltonian = TransformedHamiltonian::new(&mut math, mass_matrix, initial_kind);
        hamiltonian.set_momentum_decoherence_length(Some(self.momentum_decoherence_length));
        let switch_draw = (self.trajectory_switch_fraction * self.num_tune as f64) as u64;
        let rng = ChaCha8Rng::try_from_rng(rng).expect("Could not seed rng");
        let stats_options = self.stats_options::<M>();
        MclmcChain::new(
            math,
            hamiltonian,
            strategy,
            rng,
            chain,
            self.subsample_frequency,
            self.dynamic_step_size,
            self.trajectory_kind,
            switch_draw,
            self.max_energy_error,
            stats_options,
        )
    }

    fn hint_num_tune(&self) -> usize {
        usize_hint(self.num_tune, "num_tune")
    }

    fn hint_num_draws(&self) -> usize {
        usize_hint(self.num_draws, "num_draws")
    }

    fn num_chains(&self) -> usize {
        self.num_chains
    }

    fn seed(&self) -> u64 {
        self.seed
    }

    fn stats_options<M: Math>(&self) -> <Self::Chain<M> as SamplerStats<M>>::StatsOptions {
        StatOptions {
            adapt: GlobalStrategyStatsOptions {
                step_size: (),
                mass_matrix: (),
            },
            hamiltonian: -1,
            point: {
                let store_gradient = self.store_gradient;
                let store_unconstrained = self.store_unconstrained;
                let store_transformed = self.store_transformed;
                TransformedPointStatsOptions {
                    store_gradient,
                    store_unconstrained,
                    store_transformed,
                }
            },
            divergence: crate::dynamics::DivergenceStatsOptions {
                store_divergences: self.store_divergences,
            },
        }
    }

    fn sampler_name(&self) -> &'static str {
        "mclmc"
    }

    fn adaptation_name(&self) -> &'static str {
        "low_rank"
    }
}

impl Default for DiagNutsSettings {
    fn default() -> Self {
        default_nuts_settings(EuclideanAdaptOptions::default(), 400, 6, 1000.0)
    }
}

impl Default for LowRankNutsSettings {
    fn default() -> Self {
        let mut vals = default_nuts_settings(EuclideanAdaptOptions::default(), 800, 6, 1000.0);
        vals.adapt_options.mass_matrix_update_freq = 20;
        vals
    }
}

impl Default for FlowNutsSettings {
    fn default() -> Self {
        default_nuts_settings(FlowSettings::default(), 1500, 1, 20.0)
    }
}

type DiagNutsChain<M> = NutsChain<M, ChaCha8Rng, GlobalStrategy<M, DiagAdaptStrategy<M>>>;
type LowRankNutsChain<M> = NutsChain<M, ChaCha8Rng, GlobalStrategy<M, LowRankMassMatrixStrategy>>;

fn nuts_options(settings: &NutsSettings<impl Debug + Copy + Default + Serialize>) -> NutsOptions {
    NutsOptions {
        maxdepth: settings.maxdepth,
        mindepth: settings.mindepth,
        store_divergences: settings.store_divergences,
        check_turning: settings.check_turning,
        target_integration_time: settings.target_integration_time,
        extra_doublings: settings.extra_doublings,
        max_energy_error: settings.max_energy_error,
    }
}

impl Settings for LowRankNutsSettings {
    type Chain<M: Math> = LowRankNutsChain<M>;

    fn new_chain<M: Math, R: Rng + ?Sized>(
        &self,
        chain: u64,
        mut math: M,
        mut rng: &mut R,
    ) -> Self::Chain<M> {
        let num_tune = self.num_tune;
        let strategy = GlobalStrategy::new(&mut math, self.adapt_options, num_tune, chain);
        let mass_matrix = LowRankMassMatrix::new(&mut math, self.adapt_options.mass_matrix_options);
        let hamiltonian = TransformedHamiltonian::new(&mut math, mass_matrix, self.trajectory_kind);

        let options = nuts_options(self);

        let rng = ChaCha8Rng::try_from_rng(&mut rng).expect("Could not seed rng");

        NutsChain::new(
            math,
            hamiltonian,
            strategy,
            options,
            rng,
            chain,
            self.stats_options(),
        )
    }

    fn hint_num_tune(&self) -> usize {
        usize_hint(self.num_tune, "num_tune")
    }

    fn hint_num_draws(&self) -> usize {
        usize_hint(self.num_draws, "num_draws")
    }

    fn num_chains(&self) -> usize {
        self.num_chains
    }

    fn seed(&self) -> u64 {
        self.seed
    }

    fn stats_options<M: Math>(&self) -> <Self::Chain<M> as SamplerStats<M>>::StatsOptions {
        StatOptions {
            adapt: GlobalStrategyStatsOptions {
                mass_matrix: (),
                step_size: (),
            },
            hamiltonian: -1,
            point: {
                let store_gradient = self.store_gradient;
                let store_unconstrained = self.store_unconstrained;
                let store_transformed = self.store_transformed;
                TransformedPointStatsOptions {
                    store_gradient,
                    store_unconstrained,
                    store_transformed,
                }
            },
            divergence: crate::dynamics::DivergenceStatsOptions {
                store_divergences: self.store_divergences,
            },
        }
    }

    fn sampler_name(&self) -> &'static str {
        "nuts"
    }

    fn adaptation_name(&self) -> &'static str {
        "low_rank"
    }
}

impl Settings for DiagNutsSettings {
    type Chain<M: Math> = DiagNutsChain<M>;

    fn new_chain<M: Math, R: Rng + ?Sized>(
        &self,
        chain: u64,
        mut math: M,
        mut rng: &mut R,
    ) -> Self::Chain<M> {
        let num_tune = self.num_tune;
        let strategy = GlobalStrategy::new(&mut math, self.adapt_options, num_tune, chain);
        let mass_matrix = DiagMassMatrix::new(
            &mut math,
            self.adapt_options.mass_matrix_options.store_mass_matrix,
        );
        let potential = TransformedHamiltonian::new(&mut math, mass_matrix, self.trajectory_kind);

        let options = nuts_options(self);

        let rng = ChaCha8Rng::try_from_rng(&mut rng).expect("Could not seed rng");

        NutsChain::new(
            math,
            potential,
            strategy,
            options,
            rng,
            chain,
            self.stats_options(),
        )
    }

    fn hint_num_tune(&self) -> usize {
        usize_hint(self.num_tune, "num_tune")
    }

    fn hint_num_draws(&self) -> usize {
        usize_hint(self.num_draws, "num_draws")
    }

    fn num_chains(&self) -> usize {
        self.num_chains
    }

    fn seed(&self) -> u64 {
        self.seed
    }

    fn stats_options<M: Math>(&self) -> <Self::Chain<M> as SamplerStats<M>>::StatsOptions {
        StatOptions {
            adapt: GlobalStrategyStatsOptions {
                mass_matrix: (),
                step_size: (),
            },
            hamiltonian: -1,
            point: {
                let store_gradient = self.store_gradient;
                let store_unconstrained = self.store_unconstrained;
                let store_transformed = self.store_transformed;
                TransformedPointStatsOptions {
                    store_gradient,
                    store_unconstrained,
                    store_transformed,
                }
            },
            divergence: crate::dynamics::DivergenceStatsOptions {
                store_divergences: self.store_divergences,
            },
        }
    }

    fn sampler_name(&self) -> &'static str {
        "nuts"
    }

    fn adaptation_name(&self) -> &'static str {
        "diagonal"
    }
}

impl Settings for FlowNutsSettings {
    type Chain<M: Math> = NutsChain<M, ChaCha8Rng, ExternalTransformAdaptation>;

    fn new_chain<M: Math, R: Rng + ?Sized>(
        &self,
        chain: u64,
        mut math: M,
        mut rng: &mut R,
    ) -> Self::Chain<M> {
        let num_tune = self.num_tune;

        let strategy =
            ExternalTransformAdaptation::new(&mut math, self.adapt_options, num_tune, chain);
        let params = math
            .new_transformation(rng, math.dim(), chain)
            .expect("Failed to create external transformation");
        let transform = ExternalTransformation::new(params);
        let hamiltonian = TransformedHamiltonian::new(&mut math, transform, self.trajectory_kind);

        let options = nuts_options(self);

        let rng = ChaCha8Rng::try_from_rng(&mut rng).expect("Could not seed rng");
        NutsChain::new(
            math,
            hamiltonian,
            strategy,
            options,
            rng,
            chain,
            self.stats_options(),
        )
    }

    fn hint_num_tune(&self) -> usize {
        usize_hint(self.num_tune, "num_tune")
    }

    fn hint_num_draws(&self) -> usize {
        usize_hint(self.num_draws, "num_draws")
    }

    fn num_chains(&self) -> usize {
        self.num_chains
    }

    fn seed(&self) -> u64 {
        self.seed
    }

    fn stats_options<M: Math>(&self) -> <Self::Chain<M> as SamplerStats<M>>::StatsOptions {
        StatOptions {
            adapt: (),
            hamiltonian: (),
            point: {
                let store_gradient = self.store_gradient;
                let store_unconstrained = self.store_unconstrained;
                let store_transformed = self.store_transformed;
                TransformedPointStatsOptions {
                    store_gradient,
                    store_unconstrained,
                    store_transformed,
                }
            },
            divergence: crate::dynamics::DivergenceStatsOptions {
                store_divergences: self.store_divergences,
            },
        }
    }

    fn sampler_name(&self) -> &'static str {
        "nuts"
    }

    fn adaptation_name(&self) -> &'static str {
        "flow"
    }
}

impl Settings for FlowMclmcSettings {
    type Chain<M: Math> = crate::mclmc::MclmcChain<
        M,
        ChaCha8Rng,
        ExternalTransformAdaptation,
        ExternalTransformation<M>,
    >;

    fn new_chain<M: Math, R: Rng + ?Sized>(
        &self,
        chain: u64,
        mut math: M,
        rng: &mut R,
    ) -> Self::Chain<M> {
        use crate::dynamics::KineticEnergyKind;
        use crate::mclmc::MclmcChain;

        let num_tune = self.num_tune;
        let strategy =
            ExternalTransformAdaptation::new(&mut math, self.adapt_options, num_tune, chain);
        let params = math
            .new_transformation(rng, math.dim(), chain)
            .expect("Failed to create external transformation");
        let transform = ExternalTransformation::new(params);
        let initial_kind = match self.trajectory_kind {
            MclmcTrajectoryKind::Microcanonical => KineticEnergyKind::Microcanonical,
            MclmcTrajectoryKind::Euclidean
            | MclmcTrajectoryKind::EuclideanEarlyThenMicrocanonical => KineticEnergyKind::Euclidean,
        };
        let mut hamiltonian = TransformedHamiltonian::new(&mut math, transform, initial_kind);
        hamiltonian.set_momentum_decoherence_length(Some(self.momentum_decoherence_length));
        let switch_draw = (self.trajectory_switch_fraction * self.num_tune as f64) as u64;
        let rng = ChaCha8Rng::try_from_rng(rng).expect("Could not seed rng");
        let stats_options = self.stats_options::<M>();
        MclmcChain::new(
            math,
            hamiltonian,
            strategy,
            rng,
            chain,
            self.subsample_frequency,
            self.dynamic_step_size,
            self.trajectory_kind,
            switch_draw,
            self.max_energy_error,
            stats_options,
        )
    }

    fn hint_num_tune(&self) -> usize {
        usize_hint(self.num_tune, "num_tune")
    }

    fn hint_num_draws(&self) -> usize {
        usize_hint(self.num_draws, "num_draws")
    }

    fn num_chains(&self) -> usize {
        self.num_chains
    }

    fn seed(&self) -> u64 {
        self.seed
    }

    fn stats_options<M: Math>(&self) -> <Self::Chain<M> as SamplerStats<M>>::StatsOptions {
        StatOptions {
            adapt: (),
            hamiltonian: (),
            point: {
                let store_gradient = self.store_gradient;
                let store_unconstrained = self.store_unconstrained;
                let store_transformed = self.store_transformed;
                TransformedPointStatsOptions {
                    store_gradient,
                    store_unconstrained,
                    store_transformed,
                }
            },
            divergence: crate::dynamics::DivergenceStatsOptions {
                store_divergences: self.store_divergences,
            },
        }
    }

    fn sampler_name(&self) -> &'static str {
        "mclmc"
    }

    fn adaptation_name(&self) -> &'static str {
        "flow"
    }
}

pub fn sample_sequentially<'math, M: Math + 'math, R: Rng + ?Sized>(
    math: M,
    settings: DiagNutsSettings,
    start: &[f64],
    draws: u64,
    chain: u64,
    rng: &mut R,
) -> Result<impl Iterator<Item = Result<(Box<[f64]>, Progress)>> + 'math> {
    let mut sampler = settings.new_chain(chain, math, rng);
    sampler.set_position(start)?;
    Ok((0..draws).map(move |_| sampler.draw()))
}

#[non_exhaustive]
#[derive(Clone, Debug)]
pub struct ChainProgress {
    pub finished_draws: usize,
    pub total_draws: usize,
    pub divergences: usize,
    pub tuning: bool,
    pub started: bool,
    pub latest_num_steps: usize,
    pub total_num_steps: usize,
    pub step_size: f64,
    pub runtime: Duration,
    pub divergent_draws: Vec<usize>,
}

impl ChainProgress {
    fn new(total: usize) -> Self {
        Self {
            finished_draws: 0,
            total_draws: total,
            divergences: 0,
            tuning: true,
            started: false,
            latest_num_steps: 0,
            step_size: 0f64,
            total_num_steps: 0,
            runtime: Duration::ZERO,
            divergent_draws: Vec::new(),
        }
    }

    fn update(&mut self, stats: &Progress, draw_duration: Duration) {
        if stats.diverging & !stats.tuning {
            self.divergences += 1;
            self.divergent_draws.push(self.finished_draws);
        }
        self.finished_draws += 1;
        self.tuning = stats.tuning;

        self.latest_num_steps = stats.num_steps as usize;
        self.total_num_steps += stats.num_steps as usize;
        self.step_size = stats.step_size;
        self.runtime += draw_duration;
    }
}

#[cfg(feature = "parallel")]
enum ChainCommand {
    Resume,
    Pause,
}

#[cfg(feature = "parallel")]
struct ChainProcess<T>
where
    T: TraceStorage,
{
    stop_marker: Sender<ChainCommand>,
    trace: Arc<Mutex<Option<T::ChainStorage>>>,
    progress: Arc<Mutex<ChainProgress>>,
}

#[cfg(feature = "parallel")]
impl<T: TraceStorage> ChainProcess<T> {
    fn finalize_many(trace: T, chains: Vec<Self>) -> Result<(Option<anyhow::Error>, T::Finalized)> {
        let finalized_chain_traces = chains
            .into_iter()
            .filter_map(|chain| chain.trace.lock().expect("Poisoned lock").take())
            .map(|chain| chain.finalize())
            .collect_vec();
        trace.finalize(finalized_chain_traces)
    }

    fn progress(&self) -> ChainProgress {
        self.progress.lock().expect("Poisoned lock").clone()
    }

    fn resume(&self) -> Result<()> {
        self.stop_marker.send(ChainCommand::Resume)?;
        Ok(())
    }

    fn pause(&self) -> Result<()> {
        self.stop_marker.send(ChainCommand::Pause)?;
        Ok(())
    }

    fn start<'model, M: Model, S: Settings>(
        model: &'model M,
        chain_trace: T::ChainStorage,
        chain_id: u64,
        seed: u64,
        settings: &'model S,
        scope: &ScopeFifo<'model>,
        results: Sender<Result<()>>,
    ) -> Result<Self> {
        let (stop_marker_tx, stop_marker_rx) = channel();

        let mut rng = ChaCha8Rng::seed_from_u64(seed);
        rng.set_stream(chain_id + 1);

        let chain_trace = Arc::new(Mutex::new(Some(chain_trace)));
        let progress = Arc::new(Mutex::new(ChainProgress::new(
            settings.hint_num_draws() + settings.hint_num_tune(),
        )));

        let trace_inner = chain_trace.clone();
        let progress_inner = progress.clone();

        scope.spawn_fifo(move |_| {
            let chain_trace = trace_inner;
            let progress = progress_inner;

            let mut sample = move || {
                let logp = model
                    .math(&mut rng)
                    .context("Failed to create model density")?;
                let dim = logp.dim();

                let mut sampler = settings.new_chain(chain_id, logp, &mut rng);

                progress.lock().expect("Poisoned mutex").started = true;

                let mut initval = vec![0f64; dim];
                // TODO maxtries
                let mut error = None;
                for _ in 0..500 {
                    model
                        .init_position(&mut rng, &mut initval)
                        .context("Failed to generate a new initial position")?;
                    if let Err(err) = sampler.set_position(&initval) {
                        error = Some(err);
                        continue;
                    }
                    error = None;
                    break;
                }

                if let Some(error) = error {
                    return Err(error.context("All initialization points failed"));
                }

                let draws = settings.hint_num_tune() + settings.hint_num_draws();

                let mut msg = stop_marker_rx.try_recv();
                let mut draw = 0;
                loop {
                    match msg {
                        // The remote end is dead
                        Err(TryRecvError::Disconnected) => {
                            break;
                        }
                        Err(TryRecvError::Empty) => {}
                        Ok(ChainCommand::Pause) => {
                            msg = stop_marker_rx.recv().map_err(|e| e.into());
                            continue;
                        }
                        Ok(ChainCommand::Resume) => {}
                    }

                    let now = Instant::now();
                    let (_point, mut draw_data, mut stats, info) = sampler.expanded_draw().unwrap();

                    let mut guard = chain_trace
                        .lock()
                        .expect("Could not unlock trace lock. Poisoned mutex");

                    let Some(trace_val) = guard.as_mut() else {
                        // The trace was removed by controller thread. We can stop sampling
                        break;
                    };
                    progress
                        .lock()
                        .expect("Poisoned mutex")
                        .update(&info, now.elapsed());

                    let math = sampler.math();
                    let dims = StatsDims::from(math.deref());
                    trace_val.record_sample(
                        settings,
                        stats.get_all(&dims),
                        draw_data.get_all(math.deref()),
                        &info,
                    )?;

                    draw += 1;
                    if draw == draws {
                        break;
                    }

                    msg = stop_marker_rx.try_recv();
                }
                Ok(())
            };

            let result = sample();

            // We intentionally ignore errors here, because this means some other
            // chain already failed, and should have reported the error.
            let _ = results.send(result);
            drop(results);
        });

        Ok(Self {
            trace: chain_trace,
            stop_marker: stop_marker_tx,
            progress,
        })
    }

    fn flush(&self) -> Result<()> {
        self.trace
            .lock()
            .map_err(|_| anyhow::anyhow!("Could not lock trace mutex"))
            .context("Could not flush trace")?
            .as_mut()
            .map(|v| v.flush())
            .transpose()?;
        Ok(())
    }
}

#[cfg(feature = "parallel")]
#[derive(Debug)]
enum SamplerCommand {
    Pause,
    Continue,
    Progress,
    Flush,
    Inspect,
}

#[cfg(feature = "parallel")]
enum SamplerResponse<T: Send + 'static> {
    Ok(),
    Progress(Box<[ChainProgress]>),
    Inspect(T),
}

#[cfg(feature = "parallel")]
pub enum SamplerWaitResult<F: Send + 'static> {
    Trace(F),
    Timeout(Sampler<F>),
    Err(anyhow::Error, Option<F>),
}

#[cfg(feature = "parallel")]
pub struct Sampler<F: Send + 'static> {
    main_thread: JoinHandle<Result<(Option<anyhow::Error>, F)>>,
    commands: SyncSender<SamplerCommand>,
    responses: Receiver<SamplerResponse<(Option<anyhow::Error>, F)>>,
    results: Receiver<Result<()>>,
}

#[cfg(feature = "parallel")]
pub struct ProgressCallback {
    pub callback: Box<dyn FnMut(Duration, Box<[ChainProgress]>) + Send>,
    pub rate: Duration,
}

#[cfg(feature = "parallel")]
impl<F: Send + 'static> Sampler<F> {
    pub fn new<M, S, C, T>(
        model: M,
        settings: S,
        trace_config: C,
        num_cores: usize,
        callback: Option<ProgressCallback>,
    ) -> Result<Self>
    where
        S: Settings,
        C: StorageConfig<Storage = T>,
        M: Model,
        T: TraceStorage<Finalized = F>,
    {
        let (commands_tx, commands_rx) = sync_channel(0);
        let (responses_tx, responses_rx) = sync_channel(0);
        let (results_tx, results_rx) = channel();

        let main_thread = spawn(move || {
            let pool = ThreadPoolBuilder::new()
                .num_threads(num_cores + 1) // One more thread because the controller also uses one
                .thread_name(|i| format!("nutpie-worker-{i}"))
                .build()
                .context("Could not start thread pool")?;

            let settings_ref = &settings;
            let model_ref = &model;
            let mut callback = callback;

            pool.scope_fifo(move |scope| {
                let results = results_tx;
                let mut chains = Vec::with_capacity(settings.num_chains());

                let mut rng = ChaCha8Rng::seed_from_u64(settings.seed());
                rng.set_stream(0);

                let math = model_ref
                    .math(&mut rng)
                    .context("Could not create model density")?;
                let trace = trace_config
                    .new_trace(settings_ref, &math)
                    .context("Could not create trace object")?;
                drop(math);

                for chain_id in 0..settings.num_chains() {
                    let chain_trace_val = trace
                        .initialize_trace_for_chain(chain_id as u64)
                        .context("Failed to create trace object")?;
                    let chain = ChainProcess::start(
                        model_ref,
                        chain_trace_val,
                        chain_id as u64,
                        settings.seed(),
                        settings_ref,
                        scope,
                        results.clone(),
                    );
                    chains.push(chain);
                }
                drop(results);

                let (chains, errors): (Vec<_>, Vec<_>) = chains.into_iter().partition_result();
                if let Some(error) = errors.into_iter().next() {
                    let _ = ChainProcess::finalize_many(trace, chains);
                    return Err(error).context("Could not start chains");
                }

                let mut main_loop = || {
                    let start_time = Instant::now();
                    let mut pause_start = Instant::now();
                    let mut pause_time = Duration::ZERO;

                    let mut progress_rate = Duration::MAX;
                    if let Some(ProgressCallback { callback, rate }) = &mut callback {
                        let progress = chains.iter().map(|chain| chain.progress()).collect_vec();
                        callback(start_time.elapsed(), progress.into());
                        progress_rate = *rate;
                    }
                    let mut last_progress = Instant::now();
                    let mut is_paused = false;

                    loop {
                        let timeout = progress_rate.checked_sub(last_progress.elapsed());
                        let timeout = timeout.unwrap_or_else(|| {
                            if let Some(ProgressCallback { callback, .. }) = &mut callback {
                                let progress =
                                    chains.iter().map(|chain| chain.progress()).collect_vec();
                                let mut elapsed = start_time.elapsed().saturating_sub(pause_time);
                                if is_paused {
                                    elapsed = elapsed.saturating_sub(pause_start.elapsed());
                                }
                                callback(elapsed, progress.into());
                            }
                            last_progress = Instant::now();
                            progress_rate
                        });

                        // TODO return when all chains are done
                        match commands_rx.recv_timeout(timeout) {
                            Ok(SamplerCommand::Pause) => {
                                for chain in chains.iter() {
                                    // This failes if the thread is done.
                                    // We just want to ignore those threads.
                                    let _ = chain.pause();
                                }
                                if !is_paused {
                                    pause_start = Instant::now();
                                }
                                is_paused = true;
                                responses_tx.send(SamplerResponse::Ok()).map_err(|e| {
                                    anyhow::anyhow!(
                                        "Could not send pause response to controller thread: {e}"
                                    )
                                })?;
                            }
                            Ok(SamplerCommand::Continue) => {
                                for chain in chains.iter() {
                                    // This failes if the thread is done.
                                    // We just want to ignore those threads.
                                    let _ = chain.resume();
                                }
                                pause_time += pause_start.elapsed();
                                is_paused = false;
                                responses_tx.send(SamplerResponse::Ok()).map_err(|e| {
                                    anyhow::anyhow!(
                                        "Could not send continue response to controller thread: {e}"
                                    )
                                })?;
                            }
                            Ok(SamplerCommand::Progress) => {
                                let progress =
                                    chains.iter().map(|chain| chain.progress()).collect_vec();
                                responses_tx.send(SamplerResponse::Progress(progress.into())).map_err(|e| {
                                    anyhow::anyhow!(
                                        "Could not send progress response to controller thread: {e}"
                                    )
                                })?;
                            }
                            Ok(SamplerCommand::Inspect) => {
                                let traces = chains
                                    .iter()
                                    .filter_map(|chain| {
                                        chain
                                            .trace
                                            .lock()
                                            .expect("Poisoned lock")
                                            .as_ref()
                                            .map(|v| v.inspect())
                                    })
                                    .collect_vec();
                                let finalized_trace = trace.inspect(traces)?;
                                responses_tx.send(SamplerResponse::Inspect(finalized_trace)).map_err(|e| {
                                    anyhow::anyhow!(
                                        "Could not send inspect response to controller thread: {e}"
                                    )
                                })?;
                            }
                            Ok(SamplerCommand::Flush) => {
                                for chain in chains.iter() {
                                    chain.flush()?;
                                }
                                responses_tx.send(SamplerResponse::Ok()).map_err(|e| {
                                    anyhow::anyhow!(
                                        "Could not send flush response to controller thread: {e}"
                                    )
                                })?;
                            }
                            Err(RecvTimeoutError::Timeout) => {}
                            Err(RecvTimeoutError::Disconnected) => {
                                if let Some(ProgressCallback { callback, .. }) = &mut callback {
                                    let progress =
                                        chains.iter().map(|chain| chain.progress()).collect_vec();
                                    let mut elapsed =
                                        start_time.elapsed().saturating_sub(pause_time);
                                    if is_paused {
                                        elapsed = elapsed.saturating_sub(pause_start.elapsed());
                                    }
                                    callback(elapsed, progress.into());
                                }
                                return Ok(());
                            }
                        };
                    }
                };
                let result: Result<()> = main_loop();
                // Run finalization even if something failed
                let output = ChainProcess::finalize_many(trace, chains)?;

                result?;
                Ok(output)
            })
        });

        Ok(Self {
            main_thread,
            commands: commands_tx,
            responses: responses_rx,
            results: results_rx,
        })
    }

    pub fn pause(&mut self) -> Result<()> {
        self.commands
            .send(SamplerCommand::Pause)
            .context("Could not send pause command to controller thread")?;
        let response = self
            .responses
            .recv()
            .context("Could not recieve pause response from controller thread")?;
        let SamplerResponse::Ok() = response else {
            bail!("Got invalid response from sample controller thread");
        };
        Ok(())
    }

    pub fn resume(&mut self) -> Result<()> {
        self.commands.send(SamplerCommand::Continue)?;
        let response = self.responses.recv()?;
        let SamplerResponse::Ok() = response else {
            bail!("Got invalid response from sample controller thread");
        };
        Ok(())
    }

    pub fn flush(&mut self) -> Result<()> {
        self.commands.send(SamplerCommand::Flush)?;
        let response = self
            .responses
            .recv()
            .context("Could not recieve flush response from controller thread")?;
        let SamplerResponse::Ok() = response else {
            bail!("Got invalid response from sample controller thread");
        };
        Ok(())
    }

    pub fn inspect(&mut self) -> Result<(Option<anyhow::Error>, F)> {
        self.commands.send(SamplerCommand::Inspect)?;
        let response = self
            .responses
            .recv()
            .context("Could not recieve inspect response from controller thread")?;
        let SamplerResponse::Inspect(trace) = response else {
            bail!("Got invalid response from sample controller thread");
        };
        Ok(trace)
    }

    pub fn abort(self) -> Result<(Option<anyhow::Error>, F)> {
        drop(self.commands);
        let result = self.main_thread.join();
        match result {
            Err(payload) => std::panic::resume_unwind(payload),
            Ok(Ok(val)) => Ok(val),
            Ok(Err(err)) => Err(err),
        }
    }

    pub fn wait_timeout(self, timeout: Duration) -> SamplerWaitResult<F> {
        let start = Instant::now();
        let mut remaining = Some(timeout);
        while remaining.is_some() {
            match self.results.recv_timeout(timeout) {
                Ok(Ok(_)) => remaining = timeout.checked_sub(start.elapsed()),
                Ok(Err(e)) => return SamplerWaitResult::Err(e, None),
                Err(RecvTimeoutError::Disconnected) => match self.abort() {
                    Ok((Some(err), trace)) => return SamplerWaitResult::Err(err, Some(trace)),
                    Ok((None, trace)) => return SamplerWaitResult::Trace(trace),
                    Err(err) => return SamplerWaitResult::Err(err, None),
                },
                Err(RecvTimeoutError::Timeout) => break,
            }
        }
        SamplerWaitResult::Timeout(self)
    }

    pub fn progress(&mut self) -> Result<Box<[ChainProgress]>> {
        self.commands.send(SamplerCommand::Progress)?;
        let response = self.responses.recv()?;
        let SamplerResponse::Progress(progress) = response else {
            bail!("Got invalid response from sample controller thread");
        };
        Ok(progress)
    }
}

#[cfg(test)]
pub mod test_logps {
    use crate::{Model, math::CpuLogpFunc, math::CpuMath};
    use anyhow::Result;
    use rand::Rng;

    pub struct CpuModel<F> {
        logp: F,
    }

    impl<F> CpuModel<F> {
        pub fn new(logp: F) -> Self {
            Self { logp }
        }
    }

    impl<F> Model for CpuModel<F>
    where
        F: Send + Sync + 'static,
        for<'a> &'a F: CpuLogpFunc,
    {
        type Math<'model> = CpuMath<&'model F>;

        fn math<R: Rng + ?Sized>(&self, _rng: &mut R) -> Result<Self::Math<'_>> {
            Ok(CpuMath::new(&self.logp))
        }

        fn init_position<R: rand::prelude::Rng + ?Sized>(
            &self,
            _rng: &mut R,
            position: &mut [f64],
        ) -> Result<()> {
            position.iter_mut().for_each(|x| *x = 0.);
            Ok(())
        }
    }
}

#[cfg(test)]
mod tests {
    use crate::math::test_logps::NormalLogp;
    use crate::{
        Chain, math::CpuMath, sample_sequentially, sampler::DiagMclmcSettings,
        sampler::DiagNutsSettings, sampler::LowRankMclmcSettings, sampler::LowRankNutsSettings,
        sampler::Settings,
    };

    #[cfg(feature = "zarr")]
    use super::test_logps::CpuModel;

    use anyhow::Result;
    use itertools::Itertools;
    use pretty_assertions::assert_eq;
    use rand::{SeedableRng, rngs::StdRng};

    #[cfg(feature = "zarr")]
    use std::{
        sync::Arc,
        time::{Duration, Instant},
    };

    #[cfg(feature = "zarr")]
    use crate::{Sampler, ZarrConfig};

    #[cfg(feature = "zarr")]
    use zarrs::storage::store::MemoryStore;

    fn assert_settings_smoke<S: Settings>(settings: S) -> Result<()> {
        let logp = NormalLogp { dim: 4, mu: 0.1 };
        let math = CpuMath::new(&logp);
        let mut rng = StdRng::seed_from_u64(42);

        let stat_names = settings.stat_names(&math);
        let stat_types = settings.stat_types(&math);
        assert!(!stat_names.is_empty());
        assert_eq!(stat_names.len(), stat_types.len());

        let mut chain = settings.new_chain(0, math, &mut rng);
        chain.set_position(&vec![0.2; 4])?;
        let (_draw, _info) = chain.draw()?;
        Ok(())
    }

    #[test]
    fn all_settings_smoke() -> Result<()> {
        assert_settings_smoke(DiagNutsSettings {
            num_tune: 10,
            num_draws: 10,
            ..Default::default()
        })?;
        assert_settings_smoke(LowRankNutsSettings {
            num_tune: 10,
            num_draws: 10,
            ..Default::default()
        })?;
        assert_settings_smoke(DiagMclmcSettings {
            num_tune: 10,
            num_draws: 10,
            ..Default::default()
        })?;
        assert_settings_smoke(LowRankMclmcSettings {
            num_tune: 10,
            num_draws: 10,
            ..Default::default()
        })?;
        Ok(())
    }

    #[test]
    fn sample_chain() -> Result<()> {
        let logp = NormalLogp { dim: 10, mu: 0.1 };
        let math = CpuMath::new(&logp);
        let settings = DiagNutsSettings {
            num_tune: 100,
            num_draws: 100,
            ..Default::default()
        };
        let start = vec![0.2; 10];

        let mut rng = StdRng::seed_from_u64(42);

        let mut chain = settings.new_chain(0, math, &mut rng);

        let (_draw, info) = chain.draw()?;
        assert!(info.tuning);
        assert_eq!(info.draw, 0);

        let math = CpuMath::new(&logp);
        let chain = sample_sequentially(math, settings, &start, 200, 1, &mut rng).unwrap();
        let mut draws = chain.collect_vec();
        assert_eq!(draws.len(), 200);

        let draw0 = draws.remove(100).unwrap();
        let (vals, stats) = draw0;
        assert_eq!(vals.len(), 10);
        assert_eq!(stats.chain, 1);
        assert_eq!(stats.draw, 100);
        Ok(())
    }

    #[cfg(feature = "zarr")]
    #[test]
    fn sample_parallel() -> Result<()> {
        let logp = NormalLogp { dim: 100, mu: 0.1 };
        let settings = DiagNutsSettings {
            num_tune: 100,
            num_draws: 100,
            seed: 10,
            ..Default::default()
        };

        let model = CpuModel::new(logp.clone());
        let store = MemoryStore::new();

        let zarr_config = ZarrConfig::new(Arc::new(store));
        let mut sampler = Sampler::new(model, settings, zarr_config, 4, None)?;
        sampler.pause()?;
        sampler.pause()?;
        // TODO flush trace
        sampler.resume()?;
        let (ok, _) = sampler.abort()?;
        if let Some(err) = ok {
            Err(err)?;
        }

        let store = MemoryStore::new();
        let zarr_config = ZarrConfig::new(Arc::new(store));
        let model = CpuModel::new(logp.clone());
        let mut sampler = Sampler::new(model, settings, zarr_config, 4, None)?;
        sampler.pause()?;
        if let (Some(err), _) = sampler.abort()? {
            Err(err)?;
        }

        let store = MemoryStore::new();
        let zarr_config = ZarrConfig::new(Arc::new(store));
        let model = CpuModel::new(logp.clone());
        let start = Instant::now();
        let sampler = Sampler::new(model, settings, zarr_config, 4, None)?;

        let mut sampler = match sampler.wait_timeout(Duration::from_nanos(100)) {
            super::SamplerWaitResult::Trace(_) => {
                dbg!(start.elapsed());
                panic!("finished");
            }
            super::SamplerWaitResult::Timeout(sampler) => sampler,
            super::SamplerWaitResult::Err(_, _) => {
                panic!("error")
            }
        };

        for _ in 0..30 {
            sampler.progress()?;
        }

        match sampler.wait_timeout(Duration::from_secs(1)) {
            super::SamplerWaitResult::Trace(_) => {
                dbg!(start.elapsed());
            }
            super::SamplerWaitResult::Timeout(_) => {
                panic!("timeout")
            }
            super::SamplerWaitResult::Err(err, _) => Err(err)?,
        };

        Ok(())
    }

    #[test]
    fn sample_seq() {
        let logp = NormalLogp { dim: 10, mu: 0.1 };
        let math = CpuMath::new(&logp);
        let settings = DiagNutsSettings {
            num_tune: 100,
            num_draws: 100,
            ..Default::default()
        };
        let start = vec![0.2; 10];

        let mut rng = StdRng::seed_from_u64(42);

        let chain = sample_sequentially(math, settings, &start, 200, 1, &mut rng).unwrap();
        let mut draws = chain.collect_vec();
        assert_eq!(draws.len(), 200);

        let draw0 = draws.remove(100).unwrap();
        let (vals, stats) = draw0;
        assert_eq!(vals.len(), 10);
        assert_eq!(stats.chain, 1);
        assert_eq!(stats.draw, 100);
    }
}