init_params doesn't work since v0.30.0

[Batmaev]

It seems that the sample function ignores specified initial parameters. I provide them as sample(model, ..., init_params=...).

How to reproduce:

@model function model(y)
    mu ~ Flat()
    y ~ Normal(mu, 0.0001)
end

m = model(-94)

sample(m, NUTS(), 1, init_params = [-94])
# in v0.29.3: correctly outputs single sample with mu = -94
# in v0.30.0: outputs something random e.g. mu = -177.713

This is important because sometimes we have very narrow distribution, and the sampler may get stuck in a region with very low probability. Providing good starting point is the only way to sample something.

P.S. I use macOS and Julia 1.10.0

[devmotion]

The keyword argument is called initial_params since Turing 0.30.0/DynamicPPL 0.24/AbstractMCMC 5.

[Batmaev]

Thank you very much

[brendanwallison]

This error message is going to lead people astray (myself included):

┌ Warning: failed to find valid initial parameters in 10 tries; consider providing explicit initial parameters using the init_params keyword └ @ Turing.Inference C:\Users\brend\.julia\packages\Turing\tdsVz\src\mcmc\hmc.jl:171

[devmotion]

Seems that #2115 is problematic - it added support for AbstractMCMC 5 (without removing support for AbstractMCMC 4 even though the API is incompatible!) and DynamicPPL 0.24 without updating all internals of Turing.jl to initial_params.