Skip to content

Commit

Permalink
clear up similar functionality with JAX
Browse files Browse the repository at this point in the history
  • Loading branch information
roualdes committed May 19, 2023
1 parent 1419457 commit d8473a9
Showing 1 changed file with 13 additions and 11 deletions.
24 changes: 13 additions & 11 deletions paper/paper.md
Original file line number Diff line number Diff line change
Expand Up @@ -87,17 +87,19 @@ memory-compatible C interface makes this possible even if the host language
(e.g., R) was compiled with a different compiler, something no prior interface
which exposed Stan's log density calculations could allow.

Existing tools with similar functionalities include `JAX` [@Bradbury:2018] and
`Turing.jl` via the `JuliaAD` ecosystem [@Ge:2018]. `BridgeStan` differs from
these tools by providing access to the existing, well-known DSL for modeling
and highly efficient CPU computation of the Stan ecosystem. The Stan community
by and large uses CPU hardware, and since Stan has been tuned for CPU
performance, `BridgeStan` is more efficient than its competitors in implementing
differentiable log densities on CPUs [@Carpenter:2015; @Radul:2020;
@Tarek:2020]. Like the immutable Stan models they interface,
`BridgeStan` functions are thread safe for parallel applications. They also
admit all of the internal parallelization of Stan models, such as internal
parallel map functions and GPU-enabled matrix operations.
Existing tools with similar automatic differentiation functionality
include `JAX` [@Bradbury:2018] and `Turing.jl` via the `JuliaAD`
ecosystem [@Ge:2018]. `BridgeStan` differs from these tools by
providing access to the existing, well-known DSL for modeling and
highly efficient CPU computation of the Stan ecosystem. The Stan
community by and large uses CPU hardware, and since Stan has been
tuned for CPU performance, `BridgeStan` is more efficient than its
competitors in implementing differentiable log densities on CPUs
[@Carpenter:2015; @Radul:2020; @Tarek:2020]. Like the immutable Stan
models they interface, `BridgeStan` functions are thread safe for
parallel applications. They also admit all of the internal

This comment has been minimized.

Copy link
@bob-carpenter

bob-carpenter May 22, 2023

Collaborator

admit -> support

parallelization of Stan models, such as internal parallel map
functions and GPU-enabled matrix operations.

`BridgeStan` enables memory allocated in the host language (Julia,
Python, or R), to be reused within Stan; though any language with a C
Expand Down

0 comments on commit d8473a9

Please sign in to comment.