Experimental

Research demonstrations rather than tutorials. They explore features that are still moving and may use APIs that change between releases, so they sit outside the supported sequence. Read them for what the method can do, not as a template to copy. Each states its own caveats up front.

  • Which parts of a bursty star-formation history can you actually measure?: which parts of a bursty star formation history a fit can actually measure. tengri models burstiness as a smooth backbone times a Gaussian-process field with a PSD-governed prior, which adds a free parameter per age bin. So a fit will always return a bursty history whether or not the data said anything about it. A known history is injected and recovered to separate the measurement from the prior talking back.

  • Multi-model Bayesian model averaging on CANDELS galaxies: Bayesian model averaging over modeling choices the photometry alone does not pin down. Seven CANDELS GOODS-South galaxies at \(z \sim 1\) are fit under four configurations that vary the SFH family, stellar isochrone, dust law, and nebular treatment, then combined by evidence weight. This puts a number on how much the choice of model contributes to the error budget.

  • Nonparametric SFH fits to JWST broad and medium bands: a Prospector-style continuity SFH fit to JWST NIRCam broad plus medium band photometry at \(z = 1.5\), with individual single-galaxy and catalog posteriors timed. Bin edges log-spaced to the age of the universe set the stellar mass to within half a dex. The default fit runs four NUTS chains; the catalog runs fixed-length HMC. The page prints the wall time of each.

  • Running tengri on the Apple GPU (MPS) — experimental: running the forward model and fits on the Apple GPU via the community jax-mps backend. Needs a JAX version tengri does not pin and runs in pure float32, so it is a feasibility study rather than a recommendation. It concludes against the GPU for single-galaxy work, which is the useful part. Measured on an M4 Pro: CPU wins by 59x on a gradient and 86x on a MAP fit, batches of a few hundred reach parity, and the run-to-run spread on MPS exceeds the differences between workloads.

  • Running tengri on an NVIDIA GPU (CUDA): the same question for an NVIDIA card, where both precisions are available and nothing needs a JAX version tengri does not pin. The GPU turns out to be a width instrument rather than a fast one: measured on an RTX 3060 against a Ryzen 9 5900X, the CPU wins a single galaxy by 33x and a single MAP fit by 8.8x; the crossover sits between 128 and 512 galaxies, and a batch of 2048 photometry gradients runs 14.7x faster on the card. Two float32 results matter more than the timings: on Ampere and later, float32 matmuls silently become TF32 unless asked otherwise, which costs 4.5% on parameter error bars, and jax.grad of a raw observable returns identically zero in float32 on any device (a result confirmed by finite-difference checks that find the likelihood-path gradient wrong by about a factor of two). So float32 fitting is not safe either. The sampler section measures a fixture that is hard to sample; single-galaxy posteriors are a CPU job.

  • 7DT medium-band photometry with supplied filter curves: an instrument tengri does not carry, taken from delivered transmission curves through prediction to a fit. Three registration routes all resolve through load_filter, so a user curve works anywhere a built-in one does and can shadow it. Counts in ADU with per-band zeropoints reach an ingestible unit through AB magnitudes. Wavelength units are the part that does not announce itself: a curve tabulated in nanometers registers as a valid description of the extreme UV, lands where the SED has no flux, and the fit converges and says nothing. The range check covers that gap and cannot see microns, which is what wave_unit= settles. Worked through the 7DT set: 20 medium bands on a 25 nm grid plus g, r, i, which ship built in.