FAQ & Troubleshooting¶
Where are models downloaded to?¶
install_model(MODEL) downloads to ./models/ in the package root by default, and FlashANSR.load(directory=get_path('models', MODEL)) reads from there. Set the FLASH_ANSR_ROOT environment variable to point get_path/get_root at a different project root.
How do I run on GPU vs CPU?¶
Flash-ANSR follows standard PyTorch device handling. Move the model with .to(device) after loading:
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = FlashANSR.load(...).to(device)
If CUDA is unavailable, everything runs on CPU (slower, but functional).
Constant refinement is using too many CPU cores¶
Constant refinement runs a parallel worker pool. By default (refiner_workers=None) the pool uses every available CPU core, which oversubscribes shared machines. Cap it explicitly at load time:
Pass an explicit integer to cap the pool size (0 disables multiprocessing). This is the right knob on shared machines where the default (all cores) would oversubscribe the CPUs.
How do I reproduce v0.4.x inference behavior?¶
The v0.5 speed defaults (KV-cache, auto-batching, static decoding) are enabled by default and designed to be quality-neutral. To opt out:
from flash_ansr import SoftmaxSamplingConfig
SoftmaxSamplingConfig(choices=1024, use_cache=False, batch_size=128, static_decode=False)
Also note the v0.5 rename: the candidate-selection penalty parsimony is now length_penalty.
How do I evaluate a model and run benchmarks?¶
As of v0.6, the evaluation engine, baseline adapters, benchmarks, and metrics have moved out of flash-ansr into the companion package srbf (Symbolic Regression Benchmark Framework). Install it with pip install srbf and see the srbf repository for usage. Flash-ANSR itself still provides the FlashANSR API (.load, .fit, .predict, .compile_results) and training, but systematic evaluation and benchmarking now live in srbf.
Where do I report bugs or ask questions?¶
Open an issue at github.com/psaegert/flash-ansr/issues.