[GeophyAI]

Five self-contained commits; reviewing them one by one is easiest. Rebased on dev (#124). The CHANGELOG has the full entries.

Commits

1. feat(qp): AcousticVTI/TTI on impl='c', s_gradient modes, qP Born pair
   • CUDA kernels for the pure-qP equations of Liang et al. (2023): forward, full and boundary-saving backward (gpu / cpu / disk; fp32 / fp16 / bf16, int8 refused), orders 2-8 and runtime order. Records match eager to 1e-6 – 2e-4.
   • s_gradient='adjoint' (new default), 'frozen', 'exact'. The exact derivative of the S_n step is ill-conditioned (eager fp32 and fp64 differ by 2-3 % on the canonical gradient test).
   • AcousticVTILSRTM / AcousticTTILSRTM: the Born pair, eager and c. The compiled backward returns the mp gradient and refuses the rest.
   • PropTorch warns once when dt breaks the scheme's stability limit (the paper's eq. B-9 for VTI; a safe-side bound for TTI).
   • Docs: user-guide/qp_born.md (forward, adjoint and Born derivation), example notebook 34 (RTM / LSRTM of a tilted bed with the isotropic, VTI and TTI operators).
2. fix(acoustic): compiled backward is the exact transpose of the forward, CPML band included (Acoustic, Acoustic3D, AcousticVRZ, AcousticVRZ3D, AcousticLSRTM, AcousticLSRTM3D, ViscoAcoustic on impl='c').
   • The forward ran the bare Laplacian where db is nonzero, and the adjoints' interior bounds stopped one cell short.
   • The LSRTM adjoint applied the forward CPML operator to the adjoint field (reflectivity dot test 0.11, now 1e-6).
   • Checkpoint replays now double-buffer the CPML memory.
   • Against an eager twin of each compiled forward: records 5e-7 – 2e-5 and gradients 2e-6 – 2e-5 in every memory mode. Domain-decomposed runs stay bit-identical to one card. V100 wall time changes by -5 % to +2 %.
   • 17 docs notebooks re-run, with inversion accuracy compared before and after. The 07 / 19 helpers now read ReconState.rows.
3. fix(c): refuse gradients the compiled backward never computes
   • LSRTM / VRZ: the wavelet's gradient (it came back as zeros or wrong values).
   • LSRTM with Ckpt: the RWI vp gradient.
   • ModelParallel refuses a wavelet that requires grad.
4. fix(dd): LSRTM cpu-staged boundary race under ModelParallel
   • A per-call copy stream left a prefetch queued that overwrote a ring slot being restored. In 3-D, two identical runs differed by up to 12 % of the gradient's peak.
   • Both LSRTM backwards now use the shared BoundarySession. cpu staging is bit-identical to gpu-direct.
   • storage='disk' is refused for the pair.
5. fix(boundary): 3-D int8/fp16 restore race on edge and corner cells
   • The band-only restore wrote such cells once per band, from differently quantized copies.
   • Now one band owns each cell, so the result equals the full-grid restore bit for bit.
   • The only machine-code change is the new kernel. The 3-D int8 noise floors in gate/noise_floors.json are now null (bit-stable).

Behaviour changes

• Default eager gradient of AcousticVTI / AcousticTTI (s_gradient='adjoint'; it was 'exact').
• Records and gradients of the second-order acoustic family on impl='c' (commit 2).
• New refusals: the LSRTM / VRZ wavelet gradient and the LSRTM Ckpt vp gradient on impl='c'; a wavelet that requires grad under ModelParallel; LSRTM with disk storage under ModelParallel.

Rebase on dev

• fix(visco): per-shot gbar for batched models #124 and commit 2 touch the same visco files (visco_acoustic2d/kernels.cuh, backward.cu, visco_acoustic.py).
• They merged without conflicts and do not interact: fix(visco): per-shot gbar for batched models #124 changes the filter grids, commit 2 the history and workspace layout.
• The rebased tree equals merge(old tip, dev) exactly. Core rebuilt.

Tests

• After the rebase, on V100 with the core rebuilt: visco (fix(visco): per-shot gbar for batched models #124's batched test, CUDA and eager), slot-table and from-pool layout tests, qP — 272 passed. test_forward_scratch_from_pool.py still expected four forward scratch grids for ViscoAcoustic; commit 2 adds a fifth (the step's update), so the count is updated in that commit.
• Before the rebase, on V100: test_qp_aniso_cuda.py, test_qp_stability.py, test_equation_registry_public_api.py, test_aniso_free_surface_guard.py — 138 passed.
• When the branch was split into these commits, each intermediate state was built from scratch, imported, and passed the CPU tests.
• Not run here: the full domain-decomposition test chain.

🤖 Generated with Claude Code