Revert "Drop careless-gpu: the [cuda] pip extra install covers GPU and CPU nodes"

This reverts commit 5f25c83085.

careless-gpu/README.md

@@ -0,0 +1,28 @@
+# careless-gpu
+
+GPU build of careless (see `../careless/` for the CPU module and the full
+background). Scaling and merging of X-ray diffraction data with approximate
+Bayesian inference (variational merging), built on TensorFlow.
+
+- Upstream: https://github.com/rs-station/careless
+- License: MIT
+
+## How it differs from the careless module
+
+- CUDA is not pip-installed (no `[cuda]` extra, saves several GB): the
+  cluster `cuda/12.8.1` module provides it instead, declared as
+  `build_requires`/`runtime_deps` (same pattern as `jfjoch_viewer`).
+  `runtime_deps` matters because TensorFlow dlopens the CUDA libraries
+  when careless runs, so cuda must auto-load with this module.
+- If careless reports missing cuDNN at runtime, the cuda module does not
+  ship it; the fallback is installing `careless[cuda]` in `build` instead.
+
+The venv layout, uv usage, and cache handling are identical to `careless`.
+To share one download cache between both builds, run each with
+`UV_CACHE_DIR=<shared path> modbuild build <version>` — a cache outside
+PREFIX is kept, an in-PREFIX cache is removed after a successful install.
+
+## Adding a new version
+
+Add the version under `versions:` in `files/config.yaml`, then on a Ra
+login node: `modbuild build <version>` (modbuild/2.1.2).

careless-gpu/build

@@ -0,0 +1,26 @@
+#!/usr/bin/env modbuild
+
+pbuild::prep() {
+    :
+}
+
+pbuild::configure() {
+    # same uv pattern as the careless (CPU) module; see that build for the whys
+    UV="${UV:-$HOME/.local/bin/uv}"
+    export UV_CACHE_DIR="${UV_CACHE_DIR:-$PREFIX/.uv-cache}"
+    export UV_PYTHON_INSTALL_DIR="$PREFIX/python"
+    "$UV" venv --python 3.12 "$PREFIX/venv"
+}
+
+pbuild::compile() {
+    :
+}
+
+pbuild::install() {
+    UV="${UV:-$HOME/.local/bin/uv}"
+    export UV_CACHE_DIR="${UV_CACHE_DIR:-$PREFIX/.uv-cache}"
+    # no [cuda] extra: CUDA comes from the cuda/12.8.1 module
+    # (build_requires/runtime_deps in files/config.yaml), not from pip wheels
+    "$UV" pip install --python "$PREFIX/venv/bin/python" "careless==${V_PKG}"
+    [[ "$UV_CACHE_DIR" == "$PREFIX"* ]] && rm -rf "$UV_CACHE_DIR" || :
+}

careless-gpu/files/config.yaml

@@ -0,0 +1,13 @@
+---
+format: 1
+careless-gpu:
+  defaults:
+    group: MX
+    overlay: base
+    relstage: unstable
+  versions:
+    0.5.4:
+      config:
+        relstage: unstable
+        build_requires: ["cuda/12.8.1"]
+        runtime_deps: ["cuda/12.8.1"]

careless-gpu/modulefile

@@ -0,0 +1,16 @@
+#%Module1.0
+
+module-whatis       "careless (GPU) - merging crystallography data with variational inference"
+module-url          "https://github.com/rs-station/careless"
+module-license      "MIT"
+module-maintainer   "MX Data <jiaxin.duan@psi.ch>"
+
+module-help     "
+careless: scaling and merging of X-ray diffraction data using
+approximate Bayesian inference (TensorFlow). GPU build: CUDA is
+provided by the cuda/12.8.1 module, auto-loaded as a runtime dep.
+Use on nodes with an NVIDIA GPU; for CPU nodes load careless instead.
+"
+
+setenv          CARELESS_ENV    $PREFIX/venv
+prepend-path    PATH            $PREFIX/venv/bin

careless/README.md

@@ -11,10 +11,14 @@ inference (variational merging), built on TensorFlow.
 Upstream recommends a fresh conda env, but repo.anaconda.com is not
 reachable from Ra, so the build uses uv instead: it fetches a standalone
 Python 3.12 into `$PREFIX/python` (upstream supports >=3.9,<3.13), creates
-`$PREFIX/venv`, and runs `uv pip install careless[cuda]==<version>` into
+`$PREFIX/venv`, and runs `uv pip install careless==<version>` into it.
-it. The `[cuda]` extra ships the NVIDIA libraries (CUDA, cuDNN, ...) as
+This is the CPU build; the GPU build (with the `[cuda]` extra and the
-pip wheels, so the one install uses the GPU on GPU nodes (only the node's
+cluster `cuda/12.8.1` module) is the separate `careless-gpu` module.
-NVIDIA driver is needed — no cuda module) and falls back to CPU elsewhere.
+
+To share one uv download cache between the careless and careless-gpu
+builds, run each with `UV_CACHE_DIR=<shared path> modbuild build <version>`
+— a cache outside PREFIX is kept, an in-PREFIX cache is removed after a
+successful install.
 
 `uv` must be on PATH on the build host (e.g. `pip install --user uv`).
 

careless/build

@@ -9,8 +9,9 @@ pbuild::configure() {
     # uv fetches a standalone CPython instead; keep it inside PREFIX so the
     # venv's interpreter symlinks don't depend on the builder's home cache.
     UV="${UV:-$HOME/.local/bin/uv}"
-    # ~/.cache/uv hits the home disk quota; default the cache into PREFIX
+    # ~/.cache/uv hits the home disk quota; default the cache into PREFIX.
-    # (overridable: UV_CACHE_DIR=<path> modbuild build <version>)
+    # Overridable so careless and careless-gpu can share one cache:
+    #   UV_CACHE_DIR=<shared path> modbuild build <version>
     export UV_CACHE_DIR="${UV_CACHE_DIR:-$PREFIX/.uv-cache}"
     export UV_PYTHON_INSTALL_DIR="$PREFIX/python"
     "$UV" venv --python 3.12 "$PREFIX/venv"
@@ -23,9 +24,8 @@ pbuild::compile() {
 pbuild::install() {
     UV="${UV:-$HOME/.local/bin/uv}"
     export UV_CACHE_DIR="${UV_CACHE_DIR:-$PREFIX/.uv-cache}"
-    # [cuda] extra ships the NVIDIA stack (CUDA, cuDNN, ...) as pip wheels:
+    # CPU version; the GPU build with the [cuda] extra is the careless-gpu module
-    # uses the GPU when present (needs only the node driver), else falls back to CPU
+    "$UV" pip install --python "$PREFIX/venv/bin/python" "careless==${V_PKG}"
-    "$UV" pip install --python "$PREFIX/venv/bin/python" "careless[cuda]==${V_PKG}"
     # only auto-remove the cache if it lives inside this PREFIX (never a shared one)
     [[ "$UV_CACHE_DIR" == "$PREFIX"* ]] && rm -rf "$UV_CACHE_DIR" || :
 }

careless/modulefile

@@ -7,9 +7,8 @@ module-maintainer   "MX Data <jiaxin.duan@psi.ch>"
 
 module-help     "
 careless: scaling and merging of X-ray diffraction data using
-approximate Bayesian inference (TensorFlow). Installed with the
+approximate Bayesian inference (TensorFlow). CPU build; for
-[cuda] extra: uses NVIDIA GPUs when present (no cuda module
+NVIDIA GPU nodes load careless-gpu instead.
-needed), falls back to CPU otherwise.
 "
 
 setenv          CARELESS_ENV    $PREFIX/venv