[bug] Multinomial distribution raises RuntimeError within pyro.render_model
Author: markmcleod-mfCreated May 30, 2025Updated Aug 10, 2025
Issue Description
The multinomial distribution when passed a pyro.param as probabilities and sampled within a call of pyro.render_model raises a runtime error.
This does not occur without the render_model, or if not using a parameter, or with other distributions.
Environment
- Ubuntu 24.04,
- python 3.12
- torch 2.7.0
- pyro 1.9.1
Code Snippet
import pyro
import torch
import pyro.distributions as dist
def model():
x = pyro.param("x",torch.tensor([0.1,0.2,0.7]))
# Multinomial causes an error when within render_model
y=pyro.sample("y", dist.Multinomial(total_count=10, probs=x))
# Other distributions do not
# y=pyro.sample("y", dist.Dirichlet(concentration=x))
return y
print(model()) # tensor([1., 2., 7.]) or similar
viz = pyro.render_model(model) # Error when using Multinomial distribution
Source: pyro-ppl/pyro