release: 0.3.4

[GeophyAI]

Version bump for sweep-solver 0.3.4 (every CUDA kernel compiled once, #105): pyproject version, the CHANGELOG [0.3.4] section and compare links, the API page eyebrow.

Already on PyPI:

• Wheel 4ff13f44, 42.4 MB (0.3.3: 93.1 MB), and sdist c22da137.
• Cores cu12 67973271 and cu13 efe981db, rebuilt in manylinux_2_28. The machine code of every kernel on every arch is byte-identical to 0.3.3's (cu12 4254 SASS + 709 PTX, cu13 4963 + 709).

Checked before upload:

• This tree with the shipped cu12 core: full suite, 1895 passed, 44 skipped.
• Clean venv on an RTX 6000 Ada: the wheel loads its own core and compiles nothing. Installed tests: 543 passed, 2 skipped.
• The same wheel on an A100, in a cu13 env (torch 2.14) and a cu12 env (torch 2.5):
  • Each env loads its own core and compiles nothing.
  • Installed tests: cu13 543 passed; cu12 543 passed, 1 failed (see below).
  • Every impl='c' equation against the 0.3.3 core of the same toolkit: every array 0.3.3 reproduces bitwise is bitwise equal (3292 / 3288 + 1 rare event of the VRZ3D checkpoint family); the arrays 0.3.3 itself does not reproduce keep its spread (20 + 20 runs, also under CUDA_LAUNCH_BLOCKING=1).
  • The eager-vs-c gradient matrix logs are identical under the 0.3.3 and the 0.3.4 cu13 core.
  • LSRTM cu13 vs cu12 is within the run-to-run floor.
• The one failure: test_visco_acoustic_cuda.py::test_gradients_match_eager_toggles[False-True] on the A100 cu12 env, rel 0.0063 against a 6e-3 tolerance.
  • It is ViscoAcoustic's amplitude-only gradient. On an A100 this gradient is not reproducible run to run, and it moves with the device memory layout: holding 64 MB / 256 MB / 1 GB of device memory before the run shifts the 0.3.3 core's grad Q by 4e-4 / 3.5e-3 / 7e-4.
  • The 0.3.4 core, half the size, shifts it by a similar amount.
  • On the Ada, both cores give bitwise-identical gradients for every toggle and free-surface combination.

🤖 Generated with Claude Code