Update for Turing v0.48

[shravanngoswamii]

The interface guide described the AbstractMCMC v1 interface and its code cells were eval: false, so CI never noticed. It is rewritten on the current interface with every cell executed.

The to_distribution example is not executed because it needs BridgeStan and a Stan toolchain.

[github-actions]

Preview the changes: https://turinglang.org/docs/pr-previews/703
Please avoid using the search feature and navigation bar in PR previews!

[Copilot]

🟡 Changes recommended

Three documentation inaccuracies concerning PG resampling, discarded steps, and callback iteration semantics remain unresolved.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Updates documentation and dependencies for Turing v0.48, including a rewritten, executable AbstractMCMC interface guide.

Changes:

• Upgrades Turing and resolved dependencies.
• Modernizes sampling, threading, vectorisation, and chain documentation.
• Updates tutorials, links, and executable examples.

File summaries

File	Review
usage/vectorisation/index.qmd	Adds validation and minibatching guidance.
usage/varnamedtuple/index.qmd	Documents symbol indexing.
usage/threadsafe-evaluation/index.qmd	Describes task-local accumulators.
usage/submodels/index.qmd	Adds Stan distribution integration.
usage/sampling-options/index.qmd	Updates sampling guidance. Nit: clarify that callbacks run only for retained samples.
usage/sampler-visualisation/index.qmd	Updates sampler descriptions.
usage/mode-estimation/index.qmd	Clarifies bound validation.
usage/external-samplers/index.qmd	Documents the current external-sampler API.
tutorials/multinomial-logistic-regression/index.qmd	Fixes the NUTS API link.
tutorials/infinite-mixture-models/index.qmd	Updates random-measure imports.
tutorials/coin-flipping/index.qmd	Updates documentation links.
tutorials/bayesian-logistic-regression/index.qmd	Fixes the NUTS API link.
tutorials/bayesian-linear-regression/index.qmd	Fixes the NUTS API link.
Project.toml	Requires Turing 0.48.
Manifest.toml	Refreshes resolved dependencies.
faq/index.qmd	Revises threading guidance.
developers/inference/abstractmcmc-interface/index.qmd	Rewrites the interface guide. Nit: correct the explanation of discard_initial.
core-functionality/index.qmd	Updates sampler and chain guidance. Nit: distinguish PG’s first-sweep resampling from later conditional sweeps.
_quarto.yml	Updates the displayed version.

Review details

Suppressed comments (2)

core-functionality/index.qmd:570

• This overstates how PG uses its configured resampling scheme. In Turing 0.48 the scheme applies only to PG's first, unconditional sweep; later conditional sweeps select ancestors categorically from the weights. Please distinguish that behavior so readers do not expect StratifiedResampler (or a custom scheme) to govern every PG sweep.

By default both samplers resample with the stratified scheme whenever the effective sample size drops below half the number of particles.
The scheme and the threshold can be chosen: the scheme types are `StratifiedResampler`, `SystematicResampler` and `MultinomialResampler`, all in `Turing.Inference`, and the threshold is a fraction of the particle count.

developers/inference/abstractmcmc-interface/index.qmd:166

• discard_initial adds discarded steps before collecting N; it does not remove 2,000 entries from the requested 20,000. The current wording can make readers expect an 18,000-element result, while this call returns 20,000 transitions after running 2,000 initial discarded steps.

`sample` runs the two methods, discards the first 2000 draws, and returns the remaining transitions as a vector.

• Files reviewed: 17/19 changed files
• Comments generated: 1
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🔵 Needs a closer look

The broad documentation update and unresolved callback-accuracy nit warrant final human review.

Review details

Suppressed comments (1)

usage/sampling-options/index.qmd:278

• This overstates the callback behavior: num_warmup only selects step_warmup and does not itself discard samples, so retained warmup samples (for example, when discard_initial < num_warmup) do invoke the callback. Also, Turing v0.48's custom SMC sampling path explicitly ignores callbacks. Please document both exceptions so users do not silently miss expected callback invocations.

The `callback` keyword argument can be used to specify a function that is called each time a sample is kept.
This function should have the signature `callback(rng, model, sampler, sample, state, iteration::Int; kwargs...)`, where `state` is the sampler state that produced the sample and `iteration` counts kept samples from 1 to the number requested.
Steps that are discarded through `num_warmup`, `discard_initial` or `thinning` do not trigger the callback.

• Files reviewed: 17/19 changed files
• Comments generated: 0 new
• Review effort level: Balanced

[yebai]

I left a few minor comments above; otherwise, it looks good.

[shravanngoswamii]

@yebai Applied all your suggestions and few more fixes.