Summary

This fixes the Linux Python GPU path for ezkl.

Before this patch:

• the CLI path initialized the ICICLE backend/device through run()
• Python gen_srs, setup, and prove bypassed that path
• with ICICLE_BACKEND_INSTALL_DIR set, Python could still stay on CPU
• once Python was forced onto the CUDA path, backend loading could still fail because the ICICLE frontend shared libraries were not globally visible to later dlopen calls

After this patch:

• gen_srs, setup, and prove all initialize the GPU backend/device through the shared execution path
• Python module initialization promotes the required ICICLE frontend shared libraries to RTLD_GLOBAL on Linux GPU builds
• CUDA backend loading can resolve ICICLE symbols correctly

Root Cause

There were two linked issues.

1. Control-flow gap

• set_device() existed, but only the CLI run() path called it
• Python reached GPU-sensitive halo2 / ICICLE code directly

2. Dynamic-linking gap

• after fixing the control-flow issue, CUDA backend loading still depended on symbols exported by:
  • libicicle_device.so
  • libicicle_hash.so
  • libicicle_field_bn254.so
  • libicicle_curve_bn254.so
• in Python, those frontend libraries were not globally visible by default
• later backend dlopen calls could fail to resolve symbols like register_deviceAPI

Changes

src/execute.rs

• make set_device() idempotent with Once
• call set_device() from:
  • gen_srs_cmd
  • setup
  • prove

src/bindings/python.rs

• on Linux GPU builds, promote the ICICLE frontend shared libraries to RTLD_GLOBAL during module initialization

Local Validation

Validated locally on 2026-05-17 in WSL on an RTX 5090.

Build

cargo check --lib --no-default-features --features python-bindings,gpu-accelerated,ezkl
maturin develop --no-default-features --features python-bindings,gpu-accelerated,ezkl

Python repro

Minimal Python chain:

1. gen_settings
2. compile_circuit
3. gen_witness
4. setup
5. prove

Observed behavior:

• before patch:
  • Python logged only Registering DEVICE: device=CPU
• after only the execution-path fix:
  • Python entered CUDA setup, but backend loading still depended on global ICICLE symbols
• after the full patch:
  • Python logs Registering DEVICE: device=CUDA
  • backend loads successfully
  • setup and prove both succeed

Scope

This is intentionally narrow:

• no algorithmic proving changes
• no non-Linux Python linker changes
• no unrelated build-system or dependency changes