@mcmcjs/julia 0.13.0 → 0.14.1

package/dist/fitlib.jl

@@ -1,757 +0,0 @@
-# MCMC.js fit library, shared by the one-shot driver (driver.jl) and the
-# persistent worker (worker.jl). handle_request runs one fit or predict request
-# (Turing or JuliaBUGS) and writes the draws as the @mcmcjs/core samples
-# artifact (a flat JSON wire). This is the only place that touches the backend
-# chain type (FlexiChains), so a chain-type change is contained here.
-using JSON
-using Logging
-using Random
-using StableRNGs
-using Turing
-using Turing: VarName
-using SHA
-using Pkg
-# Imported (not `using`) so they do not pull `@model` etc. into Main and clash
-# with Turing's exports, which would break Turing model files on include.
-import JuliaBUGS
-import AdvancedHMC
-import ForwardDiff
-# Loading Mooncake activates AbstractPPL's native extension for AutoMooncake.
-import Mooncake
-# Loading ReverseDiff activates the DynamicPPL extension for AutoReverseDiff.
-import ReverseDiff
-import FlexiChains
-import DimensionalData
-
-# AbstractMCMC reports sampling progress through ProgressLogging (reachable via
-# Turing's dependency, no extra package in the managed env).
-const ProgressLogging = Turing.AbstractMCMC.ProgressLogging
-
-# Where progress JSON lines go: the driver leaves this on stderr; the worker
-# points it at the client socket for the duration of a request.
-const PROGRESS_IO = Ref{IO}(stderr)
-
-# Translates ProgressLogging records into one JSON line per update
-# ({"mcmcjs":"progress",...}). The TypeScript side filters these out of the
-# failure protocol's stderr buffer. Info-and-above records pass through to a
-# plain console logger; Debug records (AdvancedHMC step-size search) are dropped.
-struct JSONProgressLogger <: Logging.AbstractLogger
-    fallback::Logging.AbstractLogger
-    last::Dict{Int,Float64}
-end
-JSONProgressLogger() =
-    JSONProgressLogger(Logging.ConsoleLogger(stderr, Logging.Info), Dict{Int,Float64}())
-
-Logging.min_enabled_level(::JSONProgressLogger) = Logging.BelowMinLevel
-Logging.shouldlog(::JSONProgressLogger, args...) = true
-Logging.catch_exceptions(::JSONProgressLogger) = true
-
-function emit_progress(logger::JSONProgressLogger, name, fraction, done)
-    m = match(r"Chain (\d+) of (\d+)", String(name))
-    chain = m === nothing ? 1 : parse(Int, m.captures[1])
-    of = m === nothing ? 1 : parse(Int, m.captures[2])
-    f = done ? 1.0 : fraction === nothing ? 0.0 : Float64(fraction)
-    # AbstractMCMC already steps by 1%; the dedupe guards chattier emitters.
-    if done || f - get(logger.last, chain, -1.0) >= 0.01
-        logger.last[chain] = f
-        println(
-            PROGRESS_IO[],
-            JSON.json(
-                Dict(
-                    "mcmcjs" => "progress",
-                    "chain" => chain,
-                    "of" => of,
-                    "fraction" => round(f; digits = 4),
-                    "done" => done,
-                ),
-            ),
-        )
-        flush(PROGRESS_IO[])
-    end
-end
-
-function Logging.handle_message(
-    logger::JSONProgressLogger, level, message, _module, group, id, file, line; kwargs...
-)
-    progress = message isa ProgressLogging.ProgressString ? message.progress :
-        message isa ProgressLogging.Progress ? message : nothing
-    if progress !== nothing
-        emit_progress(logger, progress.name, progress.fraction, progress.done)
-        return
-    end
-    if haskey(kwargs, :progress) && (kwargs[:progress] isa Real || kwargs[:progress] == "done")
-        value = kwargs[:progress]
-        emit_progress(logger, string(message), value == "done" ? nothing : value, value == "done")
-        return
-    end
-    level >= Logging.Info || return
-    Logging.handle_message(logger.fallback, level, message, _module, group, id, file, line; kwargs...)
-end
-
-function build_adtype(name)
-    name == "forwarddiff" && return Turing.ADTypes.AutoForwardDiff()
-    name == "reversediff" && return Turing.ADTypes.AutoReverseDiff()
-    name == "mooncake" && return Turing.ADTypes.AutoMooncake(; config = nothing)
-    error("unsupported adtype: $name")
-end
-
-function build_sampler(sampler, backend)
-    algorithm = get(sampler, "algorithm", "NUTS")
-    # Prior draws are iid: no warmup, no adaptation. The JuliaBUGS path samples
-    # ancestrally (sample_bugs_prior) instead of going through a sampler object.
-    algorithm == "Prior" && return Turing.Prior(), 0
-    warmup = Int(get(sampler, "warmup", 1000))
-    adapt_delta = Float64(get(sampler, "adapt_delta", 0.8))
-    if backend == "juliabugs"
-        algorithm == "NUTS" || error("the juliabugs backend supports NUTS and Prior only")
-        return AdvancedHMC.NUTS(adapt_delta), warmup
-    end
-    adkw = haskey(sampler, "adtype") ? (; adtype = build_adtype(sampler["adtype"])) : (;)
-    algorithm == "NUTS" && return Turing.NUTS(warmup, adapt_delta; adkw...), warmup
-    algorithm == "HMC" &&
-        return Turing.HMC(Float64(sampler["step_size"]), Int(sampler["leapfrog_steps"]); adkw...),
-        warmup
-    algorithm == "HMCDA" &&
-        return Turing.HMCDA(warmup, adapt_delta, Float64(sampler["lambda"]); adkw...), warmup
-    algorithm == "MH" && return Turing.MH(), warmup
-    error("unsupported sampler algorithm: $algorithm")
-end
-
-# Extra keyword arguments for AbstractMCMC.sample, from the request's sampler
-# table: thinning, burn-in for the non-adaptive samplers (NUTS and HMCDA discard
-# their own adaptation), and named starting values replicated across chains.
-function sampling_kwargs(sampler, warmup, chains)
-    kw = (;)
-    thin = Int(get(sampler, "thin", 1))
-    thin > 1 && (kw = merge(kw, (; thinning = thin)))
-    algorithm = get(sampler, "algorithm", "NUTS")
-    algorithm in ("MH", "HMC") && warmup > 0 && (kw = merge(kw, (; discard_initial = warmup)))
-    if haskey(sampler, "initial_params")
-        nt = (; (Symbol(k) => narrow(v) for (k, v) in sampler["initial_params"])...)
-        strategy = Turing.DynamicPPL.InitFromParams(nt)
-        kw = merge(kw, (; initial_params = fill(strategy, chains)))
-    end
-    return kw
-end
-
-to_namedtuple(data) = (; (Symbol(k) => v for (k, v) in data)...)
-
-# Binds the data so a model can read a variable as a property (`data.y`) or by
-# index (`data["y"]` / `data[:y]`), with `haskey`/`keys` supported, so models
-# written in either idiom run unchanged. The underlying NamedTuple is reached
-# via getfield to avoid colliding with a data variable also named `nt`.
-struct ModelData
-    nt::NamedTuple
-end
-Base.getproperty(d::ModelData, k::Symbol) = getproperty(getfield(d, :nt), k)
-Base.getindex(d::ModelData, k::Symbol) = getproperty(getfield(d, :nt), k)
-Base.getindex(d::ModelData, k::AbstractString) = getproperty(getfield(d, :nt), Symbol(k))
-Base.haskey(d::ModelData, k::Symbol) = haskey(getfield(d, :nt), k)
-Base.haskey(d::ModelData, k::AbstractString) = haskey(getfield(d, :nt), Symbol(k))
-Base.keys(d::ModelData) = keys(getfield(d, :nt))
-Base.propertynames(d::ModelData) = propertynames(getfield(d, :nt))
-
-# JuliaBUGS requires dense arrays with a concrete numeric eltype, but JSON parses
-# them as Vector{Any} (and nested arrays as vectors of vectors). narrow promotes a
-# numeric array to a concrete eltype and stacks equal-sized nested arrays into a
-# dense array; an empty array becomes Float64[].
-function narrow(v::AbstractVector)
-    isempty(v) && return Float64[]
-    elems = [narrow(x) for x in v]
-    if all(e -> e isa AbstractArray, elems) && allequal(size.(elems))
-        return stack(elems; dims = 1)
-    end
-    return eltype(elems) <: Real && !isconcretetype(eltype(elems)) ? float.(elems) : elems
-end
-narrow(x) = x
-bugs_namedtuple(data) = (; (Symbol(k) => narrow(v) for (k, v) in data)...)
-
-# JSON null arrives as `nothing`; map it to `missing` so a blanked outcome becomes
-# a sampling (predictive) statement in the model.
-to_missing(x) = x === nothing ? missing : (x isa AbstractVector ? map(to_missing, x) : x)
-predict_namedtuple(data) = (; (Symbol(k) => to_missing(v) for (k, v) in data)...)
-
-# A blanked target compiles to a latent node; JuliaBUGS drops scalar missings.
-bugs_predict_namedtuple(data) = (; (Symbol(k) => narrow(to_missing(v)) for (k, v) in data)...)
-
-# Rebuild a posterior VNChain (latents only) from our samples wire so it can feed
-# Turing's predict. Scalar leaves are regrouped into their array-valued VarName
-# (theta[1..8] -> @varname(theta) holding an 8-vector), so predict matches the
-# model's parameters exactly instead of relying on leaf reconstruction. Target
-# outcome columns are excluded.
-function wire_to_vnchain(wire, targets)
-    sz = wire["size"]
-    nIter, total, nChains = Int(sz[1]), Int(sz[2]), Int(sz[3])
-    flat = wire["value_flat"]
-    index = Dict(n => i for (i, n) in enumerate(wire["parameters"]))
-    is_target(n) = any(t -> n == t || startswith(n, t * "["), targets)
-    keep = [n for n in wire["name_map"]["parameters"] if !is_target(n)]
-    cell(name, i, c) = begin
-        p = index[name]
-        v = flat[i + (p - 1) * nIter + (c - 1) * nIter * total]
-        v === nothing ? NaN : Float64(v)
-    end
-    base_of(n) = (b = findfirst('[', n); b === nothing ? n : n[1:prevind(n, b)])
-    idx_of(n) = begin
-        b = findfirst('[', n)
-        b === nothing && return Int[]
-        inner = n[nextind(n, b):prevind(n, findlast(']', n))]
-        parse.(Int, split(inner, ','))
-    end
-    groups = Dict{String,Vector{String}}()
-    for n in keep
-        push!(get!(groups, base_of(n), String[]), n)
-    end
-    dict = Dict{FlexiChains.ParameterOrExtra{<:VarName},Array}()
-    for (base, names) in groups
-        vn = eval(:(Turing.@varname($(Meta.parse(base)))))
-        if length(names) == 1 && isempty(idx_of(names[1]))
-            m = Matrix{Float64}(undef, nIter, nChains)
-            for c in 1:nChains, i in 1:nIter
-                m[i, c] = cell(names[1], i, c)
-            end
-            dict[FlexiChains.Parameter(vn)] = m
-        else
-            idxs = [idx_of(n) for n in names]
-            D = length(idxs[1])
-            dims = ntuple(d -> maximum(ix[d] for ix in idxs), D)
-            m = Matrix{Array{Float64,D}}(undef, nIter, nChains)
-            for c in 1:nChains, i in 1:nIter
-                a = Array{Float64,D}(undef, dims...)
-                for (n, ix) in zip(names, idxs)
-                    a[ix...] = cell(n, i, c)
-                end
-                m[i, c] = a
-            end
-            dict[FlexiChains.Parameter(vn)] = m
-        end
-    end
-    return FlexiChains.FlexiChain{VarName}(nIter, nChains, dict)
-end
-
-# Expand one sampler transition to canonical scalar leaves (theta -> theta[1], ...),
-# the same names the final samples file uses, so streamed batches reconstruct it.
-# Sampler statistics (acceptance rate, tree depth, ...) ride along under the same
-# names the samples file records as internals, so consumers get per-draw
-# diagnostics without a second channel; Bool stats become 0/1.
-function flatten_draw(transition)
-    out = Pair{String,Float64}[]
-    for (vn, val) in pairs(transition.params)
-        for (leaf, leafval) in Turing.DynamicPPL.varname_and_value_leaves(vn, val)
-            push!(out, string(leaf) => Float64(leafval))
-        end
-    end
-    if hasproperty(transition, :stats)
-        for (name, value) in pairs(transition.stats)
-            value isa Real && push!(out, string(name) => Float64(value))
-        end
-    end
-    return out
-end
-
-# A per-draw callback that emits {"mcmcjs":"draws",...} batches on PROGRESS_IO as
-# sampling proceeds. AbstractMCMC passes the 1-based chain index as `chain_number`;
-# each batch carries a per-chain monotonic seq, and a chain's batches concatenate
-# (by leaf name) to the final samples file. Assumes the callback is not invoked
-# concurrently across chains (true for the MCMCSerial path build_and_sample uses).
-function draw_streamer(draws_per_chain::Int, batch_size::Int)
-    chain = Ref(-1)
-    seq = Ref(0)
-    names = String[]
-    buffer = Dict{String,Vector{Float64}}()
-    filled = Ref(0)
-    function emit(iteration)
-        filled[] == 0 && return
-        cols = Dict{String,Vector{Float64}}(n => copy(buffer[n]) for n in names)
-        println(
-            PROGRESS_IO[],
-            JSON.json(
-                Dict(
-                    "mcmcjs" => "draws",
-                    "chain" => chain[],
-                    "seq" => seq[],
-                    "iteration" => iteration,
-                    "draws" => cols,
-                ),
-            ),
-        )
-        flush(PROGRESS_IO[])
-        seq[] += 1
-        for n in names
-            empty!(buffer[n])
-        end
-        filled[] = 0
-    end
-    return function (_rng, _model, _sampler, transition, _state, iteration; chain_number = 1, _kwargs...)
-        c = Int(chain_number) - 1
-        if c != chain[]
-            chain[] = c
-            seq[] = 0
-        end
-        leaves = flatten_draw(transition)
-        if isempty(names)
-            for (n, _) in leaves
-                push!(names, n)
-                buffer[n] = Float64[]
-            end
-        end
-        for (n, v) in leaves
-            push!(get!(buffer, n, Float64[]), v)
-        end
-        filled[] += 1
-        if filled[] >= batch_size || iteration >= draws_per_chain
-            emit(iteration)
-        end
-    end
-end
-
-# Turing observes a variable only when its name is a model argument. A dict-reading
-# model (`y = data["y"]; y[i] ~ dist`) would otherwise SAMPLE its outcome instead of
-# observing it, so condition the built model on the data columns to turn each such
-# `~` into an observation. Columns that are already model arguments are skipped: they
-# are observed by construction, and conditioning one that is also an argument errors.
-# Do not probe the model to find latent columns (e.g. via VarInfo): evaluating a
-# dict-reading model assigns prior draws into the array the data is read from,
-# corrupting it before we condition. Predictor columns carry no `~`, so conditioning
-# on them is a harmless no-op.
-function condition_on_data(model, data)
-    nt = data isa ModelData ? getfield(data, :nt) : data
-    nt isa NamedTuple || return model
-    argnames = keys(model.args)
-    pairs = [k => getproperty(nt, k) for k in keys(nt) if Base.isidentifier(string(k)) && !(k in argnames)]
-    isempty(pairs) && return model
-    return Turing.DynamicPPL.condition(model, (; pairs...))
-end
-
-# Build the model and sample in one call so both run in the latest world age:
-# the model methods are defined by `include` at runtime, so a plain call from
-# this (older) world would fail with "method too new". invokelatest fixes that.
-# Returns Turing's default chain type, a FlexiChain. With a callback, draws stream.
-function build_and_sample(entry, data, sampler, draws, chains, rng; callback = nothing, extra = (;))
-    model = condition_on_data(entry(data), data)
-    kw = callback === nothing ? extra : merge(extra, (; callback))
-    return Logging.with_logger(JSONProgressLogger()) do
-        sample(rng, model, sampler, MCMCSerial(), draws, chains; progress = true, kw...)
-    end
-end
-
-# FlexiChains-native wire writer. DimArray(chn) splits array-valued parameters
-# into scalar leaves (theta -> theta[1], theta[2]) in the (iter, chain, param)
-# orientation; Real-valued extras become the internals section.
-function vnchain_to_wire(chn)
-    da = DimensionalData.DimArray(chn)
-    pnames = string.(collect(DimensionalData.lookup(da, :param)))
-    arr = parent(da)
-    nIter, nChains, nParams = size(arr)
-
-    extras = [e for e in FlexiChains.extras(chn) if eltype(chn[e]) <: Union{Missing,Real}]
-    enames = [string(Symbol(e)) for e in extras]
-    total = nParams + length(extras)
-
-    flat = Vector{Union{Float64,Nothing}}(undef, nIter * total * nChains)
-    # JSON has no Inf/NaN; a non-finite draw (e.g. 1/sqrt(tau) under a diffuse
-    # prior) becomes null, which the samples parser reads back as NaN.
-    cell(v) = v === missing || !isfinite(v) ? nothing : Float64(v)
-    for c in 1:nChains, p in 1:nParams, i in 1:nIter
-        flat[i + (p - 1) * nIter + (c - 1) * nIter * total] = cell(arr[i, c, p])
-    end
-    for (k, e) in enumerate(extras)
-        m = chn[e]
-        p = nParams + k
-        for c in 1:nChains, i in 1:nIter
-            flat[i + (p - 1) * nIter + (c - 1) * nIter * total] = cell(m[i, c])
-        end
-    end
-
-    return Dict(
-        "size" => [nIter, total, nChains],
-        "value_flat" => flat,
-        "parameters" => vcat(pnames, enames),
-        "name_map" => Dict("parameters" => pnames, "internals" => enames),
-    )
-end
-
-# JuliaBUGS compiles the model at runtime, so building and sampling must run in
-# separate latest-world frames; merging them reintroduces a world-age error.
-# Returns a FlexiChain like the Turing path; gen_chains recovers generated
-# quantities (unobserved nodes with no observed descendants) after sampling.
-# With a callback, draws stream (see bugs_draw_streamer).
-function sample_bugs(model, sampler, warmup, draws, chains, rng; callback = nothing, thin = 1)
-    kw = callback === nothing ? (;) : (; callback)
-    thin > 1 && (kw = merge(kw, (; thinning = thin)))
-    return Logging.with_logger(JSONProgressLogger()) do
-        JuliaBUGS.AbstractMCMC.sample(
-            rng, model, sampler, JuliaBUGS.AbstractMCMC.MCMCSerial(), draws, chains;
-            chain_type = FlexiChains.VNChain, n_adapts = warmup, discard_initial = warmup,
-            progress = true, kw...,
-        )
-    end
-end
-
-# JuliaBUGS prior sampling: each draw is one ancestral pass (evaluate!! with an
-# rng samples every unobserved stochastic node from its prior and recomputes the
-# deterministic ones; observed data stays fixed). The collected variables are the
-# model parameters plus generated quantities, the same columns a posterior fit
-# produces through gen_chains.
-function sample_bugs_prior(model, draws, chains, rng)
-    model isa JuliaBUGS.BUGSModelWithGradient && (model = model.base_model)
-    APPL = JuliaBUGS.AbstractPPL
-    vars = vcat(JuliaBUGS.model_parameters(model), JuliaBUGS.generated_quantities(model))
-    isempty(vars) && error("the model has no parameters or generated quantities to sample")
-    grab(env, vn) = (v = APPL.getvalue(env, vn); v isa AbstractArray ? copy(v) : v)
-    columns = Dict(vn => Vector{Any}(undef, draws * chains) for vn in vars)
-    for c in 1:chains, i in 1:draws
-        env, _ = APPL.evaluate!!(rng, model)
-        for vn in vars
-            columns[vn][i + (c - 1) * draws] = grab(env, vn)
-        end
-    end
-    dict = Dict{FlexiChains.ParameterOrExtra{<:VarName},Array}(
-        FlexiChains.Parameter(vn) =>
-            [columns[vn][i + (c - 1) * draws] for i in 1:draws, c in 1:chains] for vn in vars
-    )
-    return FlexiChains.FlexiChain{VarName}(draws, chains, dict)
-end
-
-# The JuliaBUGS counterpart to draw_streamer. A JuliaBUGS transition is a raw
-# AdvancedHMC.Transition whose params are an unconstrained vector, so (unlike the
-# Turing path) draws cannot be read off it directly. Per batch we rebuild a small
-# FlexiChain through the same gen_chains + vnchain_to_wire path the final file
-# uses, so the streamed leaf names (parameters, generated quantities, and sampler
-# stats) match the file exactly. Reconstruction runs on a copy of the model so it
-# never perturbs the sampler's state. AbstractMCMC passes the 1-based chain index
-# as chain_number; each chain's batches carry a per-chain monotonic seq.
-# TODO: once the pinned JuliaBUGS carries TuringLang/JuliaBUGS.jl#517 (with
-# AbstractMCMC#212 and AbstractPPL#178), a transition exposes named constrained
-# params (a VarNamedTuple) plus generated quantities, so this collapses to a
-# per-draw flatten like draw_streamer, dropping the deepcopy and reconstruction.
-function bugs_draw_streamer(model, draws_per_chain::Int, batch_size::Int)
-    recon = deepcopy(model)
-    side_rng = StableRNG(0)
-    chain = Ref(-1)
-    seq = Ref(0)
-    θs = Vector{Vector{Float64}}()
-    lps = Float64[]
-    stats = Any[]
-    function emit(iteration)
-        isempty(θs) && return
-        stats_names = collect(keys(merge((; lp = lps[1]), stats[1])))
-        stats_values = [vcat(lps[i], collect(values(stats[i]))) for i in eachindex(θs)]
-        chn = JuliaBUGS.gen_chains(FlexiChains.VNChain, recon, θs, stats_names, stats_values; rng = side_rng)
-        wire = vnchain_to_wire(chn)
-        nIter, total, _ = wire["size"]
-        flat = wire["value_flat"]
-        params = wire["parameters"]
-        cols = Dict(params[p] => [flat[i + (p - 1) * nIter] for i in 1:nIter] for p in 1:total)
-        println(
-            PROGRESS_IO[],
-            JSON.json(
-                Dict(
-                    "mcmcjs" => "draws",
-                    "chain" => chain[],
-                    "seq" => seq[],
-                    "iteration" => iteration,
-                    "draws" => cols,
-                ),
-            ),
-        )
-        flush(PROGRESS_IO[])
-        seq[] += 1
-        empty!(θs)
-        empty!(lps)
-        empty!(stats)
-    end
-    return function (_rng, _model, _sampler, transition, _state, iteration; chain_number = 1, _kwargs...)
-        c = Int(chain_number) - 1
-        if c != chain[]
-            chain[] = c
-            seq[] = 0
-            empty!(θs)
-            empty!(lps)
-            empty!(stats)
-        end
-        push!(θs, copy(transition.z.θ))
-        push!(lps, transition.z.ℓπ.value)
-        push!(stats, AdvancedHMC.stat(transition))
-        if length(θs) >= batch_size || iteration >= draws_per_chain
-            emit(iteration)
-        end
-    end
-end
-
-# JuliaBUGS pointwise log-likelihood: restore each posterior draw into the
-# evaluation environment, recompute the deterministic nodes in topological
-# order, and record logpdf per observed node. Every unobserved stochastic node
-# must be covered by the posterior columns, or log p(y | theta) would not be a
-# function of the draw.
-function loglik_bugs(model, posterior)
-    model isa JuliaBUGS.BUGSModelWithGradient && (model = model.base_model)
-    APPL = JuliaBUGS.AbstractPPL
-    gd = model.graph_evaluation_data
-
-    sz = posterior["size"]
-    nIter, nCols, nChains = Int(sz[1]), Int(sz[2]), Int(sz[3])
-    flat = posterior["value_flat"]
-    col = Dict(n => i for (i, n) in enumerate(posterior["parameters"]))
-    posterior_names = Set(posterior["name_map"]["parameters"])
-    cell(i, p, c) = begin
-        v = flat[i + (p - 1) * nIter + (c - 1) * nIter * nCols]
-        v === nothing ? NaN : Float64(v)
-    end
-
-    restore = Tuple{APPL.VarName,Int}[]
-    for vn in JuliaBUGS.model_parameters(model)
-        leaves = collect(APPL.varname_leaves(vn, APPL.getvalue(model.evaluation_env, vn)))
-        found = count(l -> string(l) in posterior_names, leaves)
-        found == length(leaves) ||
-            error("the posterior samples do not cover $(vn); were they produced by fitting this spec?")
-        append!(restore, [(l, col[string(l)]) for l in leaves])
-    end
-    observed = [
-        (vn, i) for (i, vn) in enumerate(gd.sorted_nodes) if
-        gd.is_stochastic_vals[i] && gd.is_observed_vals[i]
-    ]
-    isempty(observed) && error("the model has no observed variables to compute a log-likelihood for")
-
-    pnames = string.(first.(observed))
-    n = length(pnames)
-    flat_out = Vector{Union{Float64,Nothing}}(undef, nIter * n * nChains)
-    for c in 1:nChains, i in 1:nIter
-        env = model.evaluation_env
-        for (leaf, p) in restore
-            env = JuliaBUGS.BangBang.setindex!!(env, cell(i, p, c), leaf)
-        end
-        for (k, node) in enumerate(gd.sorted_nodes)
-            if !gd.is_stochastic_vals[k]
-                env = JuliaBUGS.BangBang.setindex!!(
-                    env, gd.node_function_vals[k](env, gd.loop_vars_vals[k]), node,
-                )
-            end
-        end
-        for (k, (vn, idx)) in enumerate(observed)
-            dist = gd.node_function_vals[idx](env, gd.loop_vars_vals[idx])
-            v = logpdf(dist, APPL.getvalue(env, vn))
-            flat_out[i + (k - 1) * nIter + (c - 1) * nIter * n] = isfinite(v) ? Float64(v) : nothing
-        end
-    end
-    return Dict(
-        "size" => [nIter, n, nChains],
-        "value_flat" => flat_out,
-        "parameters" => pnames,
-        "name_map" => Dict("parameters" => pnames, "internals" => String[]),
-    )
-end
-
-# JuliaBUGS posterior prediction: condition the (target-blanked) model on the
-# posterior parameter columns, then evaluate!! ancestral-samples the remaining
-# latents once per posterior draw. No sampler or gradient runs.
-function predict_bugs(model, posterior, targets, rng)
-    model isa JuliaBUGS.BUGSModelWithGradient && (model = model.base_model)
-    APPL = JuliaBUGS.AbstractPPL
-    target_syms = Set(Symbol.(targets))
-    is_target(vn) = APPL.getsym(vn) in target_syms
-
-    sz = posterior["size"]
-    nIter, nCols, nChains = Int(sz[1]), Int(sz[2]), Int(sz[3])
-    flat = posterior["value_flat"]
-    col = Dict(n => i for (i, n) in enumerate(posterior["parameters"]))
-    posterior_names = Set(posterior["name_map"]["parameters"])
-    cell(i, p, c) = begin
-        v = flat[i + (p - 1) * nIter + (c - 1) * nIter * nCols]
-        v === nothing ? NaN : Float64(v)
-    end
-
-    # Nodes fully covered by posterior columns are conditioned; anything
-    # uncovered stays latent and is forward-sampled, like Turing.predict.
-    gd = model.graph_evaluation_data
-    conditioned = APPL.VarName[]
-    restore = Tuple{APPL.VarName,Int}[]
-    for (i, vn) in enumerate(gd.sorted_nodes)
-        (gd.is_stochastic_vals[i] && !gd.is_observed_vals[i] && !is_target(vn)) || continue
-        leaves = collect(APPL.varname_leaves(vn, APPL.getvalue(model.evaluation_env, vn)))
-        found = count(l -> string(l) in posterior_names, leaves)
-        if found == length(leaves)
-            push!(conditioned, vn)
-            append!(restore, [(l, col[string(l)]) for l in leaves])
-        elseif found > 0
-            error("the posterior samples cover $(vn) only partially; refit before predicting")
-        end
-    end
-    isempty(conditioned) &&
-        error("the posterior samples share no parameters with the model; were they produced by fitting this spec?")
-    cond = APPL.condition(model, conditioned)
-
-    any(is_target, cond.graph_evaluation_data.sorted_nodes) ||
-        error("no variables predicted; check that [predict].targets name model variables")
-
-    out_leaves = APPL.VarName[]
-    arr = Array{Float64,3}(undef, 0, 0, 0)
-    for c in 1:nChains, i in 1:nIter
-        env = cond.evaluation_env
-        for (leaf, p) in restore
-            env = JuliaBUGS.BangBang.setindex!!(env, cell(i, p, c), leaf)
-        end
-        cond = JuliaBUGS.BangBang.setproperty!!(cond, :evaluation_env, env)
-        drawn, _ = APPL.evaluate!!(rng, cond)
-        if isempty(out_leaves)
-            out_leaves = [
-                l for vn in cond.graph_evaluation_data.sorted_nodes if is_target(vn)
-                for l in APPL.varname_leaves(vn, APPL.getvalue(drawn, vn))
-            ]
-            arr = Array{Float64,3}(undef, nIter, length(out_leaves), nChains)
-        end
-        for (k, leaf) in enumerate(out_leaves)
-            arr[i, k, c] = Float64(APPL.getvalue(drawn, leaf))
-        end
-    end
-
-    pnames = string.(out_leaves)
-    n = length(pnames)
-    flat_out = Vector{Union{Float64,Nothing}}(undef, nIter * n * nChains)
-    for c in 1:nChains, p in 1:n, i in 1:nIter
-        v = arr[i, p, c]
-        flat_out[i + (p - 1) * nIter + (c - 1) * nIter * n] = isfinite(v) ? v : nothing
-    end
-    return Dict(
-        "size" => [nIter, n, nChains],
-        "value_flat" => flat_out,
-        "parameters" => pnames,
-        "name_map" => Dict("parameters" => pnames, "internals" => String[]),
-    )
-end
-
-function provenance()
-    packages = Dict{String,String}()
-    for (_, info) in Pkg.dependencies()
-        if info.name in
-           ("Turing", "FlexiChains", "DynamicPPL", "AbstractMCMC", "JuliaBUGS", "AdvancedHMC", "ForwardDiff", "Mooncake", "ADTypes") &&
-           info.version !== nothing
-            packages[info.name] = string(info.version)
-        end
-    end
-    manifest = joinpath(dirname(Base.active_project()), "Manifest.toml")
-    return Dict(
-        "julia_version" => string(VERSION),
-        "packages" => packages,
-        "manifest_sha256" => isfile(manifest) ? bytes2hex(sha256(read(manifest))) : "",
-    )
-end
-
-# Resolve the model entry function: use the requested name, falling back to the
-# conventional build_model. This lets a request leave the entry implicit, or name a
-# custom entry that a given model file does not define (then build_model is used).
-# Load the user's model file into a throwaway module. Isolation keeps repeated
-# requests in the persistent worker from colliding on names (e.g. two models each
-# defining `const model_def`), and confines the model's globals to their own scope.
-function load_model_module(path)
-    mod = Module(gensym(:UserModel))
-    Base.include(mod, abspath(path))
-    return mod
-end
-
-function resolve_entry(mod, requested)
-    for name in (requested, "build_model")
-        name === nothing && continue
-        sym = Symbol(name)
-        # The model was just include'd, so its bindings live in a newer world than
-        # this frame; resolve them in the latest world to avoid a Julia 1.12
-        # world-age warning on the access.
-        Base.invokelatest(isdefined, mod, sym) && return Base.invokelatest(getfield, mod, sym)
-    end
-    error("model file defines no entry function (looked for $(requested) and build_model)")
-end
-
-# Runs one request end to end. Returns {"ok": true, "provenance": ...} or
-# {"ok": false, "error": ..., "stage": ...}; never throws.
-function handle_request(request)
-    out = request["out"]
-    mode = get(request, "mode", "fit")
-    backend = get(get(request, "backend", Dict()), "id", "turing")
-    stage = "compile"
-    try
-        Random.seed!(Int(request["seed"]))
-        rng = StableRNG(Int(request["seed"]))
-        modelmod = load_model_module(request["model"]["file"])
-        entry = resolve_entry(modelmod, get(request["model"], "entry", nothing))
-
-        wire = if mode == "loglik"
-            if backend == "juliabugs"
-                model = Base.invokelatest(entry, bugs_namedtuple(request["data"]))
-                stage = "load_samples"
-                posterior = JSON.parsefile(request["samples"])
-                stage = "loglik"
-                Base.invokelatest(loglik_bugs, model, posterior)
-            else
-                data = ModelData(to_namedtuple(request["data"]))
-                model = Base.invokelatest(entry, data)
-                model = Base.invokelatest(condition_on_data, model, data)
-                stage = "load_samples"
-                chn = wire_to_vnchain(JSON.parsefile(request["samples"]), String[])
-                stage = "loglik"
-                pll = Base.invokelatest(Turing.DynamicPPL.pointwise_loglikelihoods, model, chn)
-                isempty(FlexiChains.parameters(pll)) &&
-                    error("the model has no observed variables to compute a log-likelihood for")
-                vnchain_to_wire(pll)
-            end
-        elseif mode == "predict"
-            if backend == "juliabugs"
-                model = Base.invokelatest(entry, bugs_predict_namedtuple(request["data"]))
-                stage = "load_samples"
-                posterior = JSON.parsefile(request["samples"])
-                stage = "predict"
-                Base.invokelatest(predict_bugs, model, posterior, request["predict"]["targets"], rng)
-            else
-                model = Base.invokelatest(entry, ModelData(predict_namedtuple(request["data"])))
-                stage = "load_samples"
-                rchn = wire_to_vnchain(JSON.parsefile(request["samples"]), request["predict"]["targets"])
-                stage = "predict"
-                pp = Base.invokelatest(predict, rng, model, rchn; include_all = false)
-                isempty(FlexiChains.parameters(pp)) &&
-                    error("no variables predicted; check that [predict].targets name unconditioned outcomes")
-                vnchain_to_wire(pp)
-            end
-        else
-            data = backend == "juliabugs" ? bugs_namedtuple(request["data"]) : ModelData(to_namedtuple(request["data"]))
-            sampler, warmup = build_sampler(request["sampler"], backend)
-            draws = Int(request["sampler"]["draws"])
-            chains = Int(get(request["sampler"], "chains", 4))
-            stage = "sample"
-            if backend == "juliabugs"
-                model = Base.invokelatest(entry, data)
-                chn = if get(request["sampler"], "algorithm", "NUTS") == "Prior"
-                    Base.invokelatest(sample_bugs_prior, model, draws, chains, rng)
-                else
-                    # Reconstruction reads the plain BUGSModel; entry may return a gradient wrapper.
-                    base = hasproperty(model, :base_model) ? model.base_model : model
-                    cb = get(request, "stream_draws", false) ?
-                        bugs_draw_streamer(base, draws, Int(get(request, "draw_batch_size", 25))) : nothing
-                    Base.invokelatest(
-                        sample_bugs, model, sampler, warmup, draws, chains, rng;
-                        callback = cb, thin = Int(get(request["sampler"], "thin", 1)),
-                    )
-                end
-                vnchain_to_wire(chn)
-            else
-                cb = get(request, "stream_draws", false) ?
-                    draw_streamer(draws, Int(get(request, "draw_batch_size", 25))) : nothing
-                extra = sampling_kwargs(request["sampler"], warmup, chains)
-                chn = Base.invokelatest(
-                    build_and_sample, entry, data, sampler, draws, chains, rng;
-                    callback = cb, extra,
-                )
-                vnchain_to_wire(chn)
-            end
-        end
-
-        stage = "write"
-        tmp = out * ".tmp"
-        open(tmp, "w") do io
-            JSON.print(io, wire)
-        end
-        mv(tmp, out; force = true)
-
-        return Dict("ok" => true, "provenance" => provenance())
-    catch err
-        return Dict("ok" => false, "error" => sprint(showerror, err), "stage" => stage)
-    end
-end