[Sophia Wen (hfwen0502)]

Fixes #73.

Build. Every install now builds the CPU backend, the unrestricted (-D_UHF) builds of the CPU and Thrust backends, and every GPU backend the detected toolchain supports: Thrust and OpenMP offload under nvc++, OpenMP offload under amdclang++.

• New SBD_SKIP_BACKENDS leaves GPU backends out, as a comma-separated list of gpu (alias gpu_thrust) and gpu_omp_offload. It refuses cpu and unknown names.
• A summary at the end of the build lists each backend, and for a GPU backend it did not build, why: skipped, no GPU compiler found, or NVIDIA only.
• Why the CPU backend is always built: it is the default device, and it carries the variance evaluator the Thrust backend lacks. A build without it failed a solve with no device=, and the test suite unless SBD_TEST_DEVICE was set (SBD_TEST_DEVICE must be set to test a build without the CPU backend #73). With the UHF builds always present, a session can take restricted and unrestricted FCIDUMPs alike.
• SBD_BUILD_BACKEND is deprecated. It still works, with a warning, and is read as a skip list: cpu skips both GPU backends, gpu the offload one, gpu_omp_offload the Thrust one; auto and all skip nothing. Each now also builds the CPU backend. SBD_BUILD_BACKEND=all no longer stops the build when no GPU compiler is found; the summary says only the CPU backend was built.
• SBD_BUILD_UHF is no longer read; setting it warns.

CPU-only analyses take no device. sbd.analysis.variance() and caop_variance() always run upstream's evaluator on the CPU, with MPI and OpenMP, whichever device the state was solved on, and sbd.spectrum.green_function_from_coefficients() reevaluates saved coefficients on the CPU. They now use the CPU backend, except that variance, in a process that has already loaded the OpenMP-offload backend, calls the same CPU code from that module, since loading the CPU backend next to it would leave offload running on the host for the rest of the process. The device argument, which only chose between modules running the same CPU code, is removed from all three; the release note lists this under upgrade notes. green_function() still takes cpu or gpu, and refuses gpu-omp.

Error messages name only devices. When a requested backend was missing, the errors from init() and get_backend() listed the internal unrestricted builds (cpu-uhf, gpu-uhf) among the usable devices. They now list what sbd.available_backends() does. When no device was given, init() says the default cpu backend is the one missing.

Docs.

• The README's backend section is now a feature-by-backend table: TPB, GDB, the CAOP ground state and Green's function, UHF FCIDUMPs, and variance and PT2.
• INSTALL.md documents SBD_SKIP_BACKENDS and the deprecations, and both it and the docs front page point to the README table.
• Runtime error messages no longer suggest SBD_BUILD_BACKEND or SBD_BUILD_UHF; a missing CPU backend now says the install is incomplete.
• CI's docs job and tox.ini use SBD_SKIP_BACKENDS.

A test fix found along the way. The in-process CAOP tests built their config from the backend under test but solved with no device, so they used the session's default device, which depends on which test initializes the session first. Run after the spectrum tests, the Thrust-only check that t_comm_size > b_comm_size is refused ran on the CPU instead and failed. Each solve now names the backend its config came from.

Checked:

• Each variable case prints the expected plan or error.
• On an 8× H100 server with NVHPC:
  • each skip combination and the deprecated SBD_BUILD_BACKEND=gpu_omp_offload build exactly the expected modules;
  • SBD_TEST_DEVICE must be set to test a build without the CPU backend #73's reproducer, a build that previously had no CPU backend tested with SBD_TEST_DEVICE unset, now passes;
  • the full suite passes on the cpu, gpu and gpu-omp backends, single-process and under MPI;
  • after a gpu-omp solve, variance runs from the offload module without loading the CPU backend;
  • the CAOP and spectrum tests pass in either order.
• On an 8× MI250X server with ROCm:
  • the default build makes the CPU and OpenMP-offload backends and reports Thrust as NVIDIA only; SBD_SKIP_BACKENDS=gpu_omp_offload builds the CPU backend alone;
  • the suite passes on the cpu and gpu-omp backends, single-process and under MPI;
  • after a gpu-omp solve, variance runs from the offload module without loading the CPU backend;
  • requesting --device gpu names only cpu and gpu-omp as usable.
• The docs build without warnings.

🤖 Generated with Claude Code