In-place operations extended in API.

.github/dependabot.yml

@@ -5,3 +5,7 @@ updates:
     directory: "/" # Location of package manifests
     schedule:
       interval: "monthly"
+  - package-ecosystem: "pip"
+    directory: "/" # root requirements.txt (docs/requirements.txt is intentionally pinned, left out)
+    schedule:
+      interval: "monthly"

.github/workflows/CI.yml

@@ -10,6 +10,10 @@ on:
 permissions:
   contents: read
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   test:
     name: Test on ${{ matrix.os }} / Python ${{ matrix.python-version }}
@@ -57,7 +61,7 @@ jobs:
       # 7. Run tests with coverage
       - name: Run tests
         run: |
-          pytest tests --runslow --cov=dgamore --cov-report=term-missing --cov-report=xml -vv
+          pytest tests --runslow --cov=dgamore --cov-report=term-missing --cov-report=xml --cov-fail-under=85 -vv
         # pytest tests --runslow --cov=dgamore --cov-report=xml --cov-report=term
 
       # 8. Upload coverage to Codecov (requires CODECOV_TOKEN for private repos)

.github/workflows/Docs.yml

@@ -0,0 +1,42 @@
+name: Docs
+
+on:
+  push:
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  build:
+    name: Build docs (Sphinx)
+    runs-on: ubuntu-latest
+    steps:
+      # 1. Checkout repository
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      # 2. Set up Python (mirrors the Read the Docs build: Python 3.13)
+      - name: Set up Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.13"
+          cache: 'pip'
+
+      # 3. Install documentation dependencies (mpi4py and cupy are mocked in conf.py)
+      - name: Install documentation dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r docs/requirements.txt
+
+      # 4. Build HTML docs; -W turns warnings into errors, --keep-going reports them all
+      - name: Build HTML documentation
+        run: |
+          sphinx-build -W --keep-going -b html docs docs/_build/html

.github/workflows/Lint.yml

@@ -0,0 +1,38 @@
+name: Lint
+
+on:
+  push:
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  black:
+    name: Black format check
+    runs-on: ubuntu-latest
+    steps:
+      # 1. Checkout repository
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      # 2. Set up Python
+      - name: Set up Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.12"
+
+      # 3. Install Black (pinned; keep local Black matching this version)
+      - name: Install Black
+        run: pip install black==26.5.1
+
+      # 4. Verify formatting (line-length also read from [tool.black] in pyproject.toml)
+      - name: Check formatting
+        run: black --check --line-length 120 .

CONTRIBUTING.md

@@ -69,16 +69,18 @@ When you would like to contribute code, the following workflow keeps things smoo
    ```bash
    pytest tests                       # fast suite (skips tests marked slow)
    pytest tests --runslow             # full suite, as run in CI
-   pytest tests --runslow --cov=dgamore --cov-report=term-missing   # with coverage
+   pytest tests --runslow --cov=dgamore --cov-report=term-missing --cov-fail-under=85   # coverage, as CI runs it
    ```
 5. **Open a pull request** against the `main` branch, with a short description of what you changed and why. If your
    pull request is related to an existing issue, mentioning it helps connect the two.
 
-A continuous integration pipeline runs the full test suite on every pull request, across Python 3.12 to 3.14 on both
-Linux and macOS. This is there to catch regressions, not to be a gatekeeper, so please do not worry if something turns
-red on the first try; it is a normal part of the process, and we are glad to help you get it passing. A coverage tool
-also checks that the overall test coverage stays above 85%, so adding tests for your changes is the best way to keep it
-healthy.
+A continuous integration pipeline runs on every pull request. It checks that the code is Black-formatted, then runs the
+full test suite across Python 3.12 to 3.14 on both Linux and macOS. This is there to catch regressions, not to be a
+gatekeeper, so please do not worry if something turns red on the first try; it is a normal part of the process, and we
+are glad to help you get it passing. The pipeline also requires the overall test coverage to stay at **at least 85%**,
+and the build fails if it drops below that threshold. Beyond the overall figure, the new or changed code in a pull
+request (the *patch*) must itself be covered to **at least 85%**, so please add tests for what you write rather than
+relying on the rest of the code base to carry the average.
 
 ## Coding style
 

README.md

@@ -49,7 +49,7 @@ Configure a run by editing your configuration file, then execute the routine wit
 holding it and `-c` to name it (defaults: the repository directory and [dga_config.yaml](dgamore/dga_config.yaml)):
 
 ```bash
-mpiexec -np 8 DGAmore.py -p /configs/ -c my_config.yaml   # or: DGAmore.py for a single-core test run
+mpiexec -np 8 DGAmore -p /configs/ -c my_config.yaml   # or: DGAmore for a single-core test run
 ```
 
 See the [installation](https://dgamore.readthedocs.io/en/latest/installation.html) and

codecov.yml

@@ -26,7 +26,7 @@ coverage:
       default:
         target: 85%
         threshold: 0%
-        informational: true
+        informational: false
 comment:
   layout: "reach, diff, flags, files"
   behavior: default

dgamore/DGAmore.py

@@ -2,11 +2,11 @@
 # SPDX-FileCopyrightText: 2025-2026 Julian Peil <julian.peil@tuwien.ac.at>
 # SPDX-License-Identifier: MIT
 #
-# DGAmore — Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
+# DGAmore - Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
 #           Eliashberg Equation Solver for Strongly Correlated Electron Systems
 """
-Main entry point and top-level driver of a DGAmore run (installed on the PATH as ``DGAmore.py``). The
-:func:`execute_dga_routine` orchestrates the full pipeline: parse the config, load the DMFT input, run the local
+Main entry point and top-level driver of a DGAmore run (installed on the PATH as ``DGAmore``). The
+:func:`main` orchestrates the full pipeline: parse the config, load the DMFT input, run the local
 Schwinger-Dyson step per inequivalent atom and assemble the full multi-band quantities, run the non-local ladder
 DGA self-energy, optionally analytically continue to real frequencies, and optionally solve the Eliashberg
 equation -- saving and plotting results along the way. Rank 0 owns the file I/O, local assembly and plotting; the
@@ -16,7 +16,11 @@
 import itertools as it
 import logging
 import os
-from copy import deepcopy
+
+# OpenMPI: exclude the UCX one-sided (RMA) component before MPI is initialised. On some OpenMPI 5.x builds it fails its
+# own component-query and prints a benign "OSC UCX component priority set inside component query failed" warning when
+# the per-node shared-memory giwk window is created.
+os.environ.setdefault("OMPI_MCA_osc", "^ucx")
 
 import matplotlib.pyplot as plt
 import numpy as np
@@ -42,7 +46,7 @@
 logging.getLogger("matplotlib").setLevel(logging.WARNING)
 
 
-def execute_dga_routine():
+def main():
     """
     Runs the complete DGA pipeline end to end: config parsing and folder setup, DMFT input loading, the local
     Schwinger-Dyson step (per inequivalent atom, assembled into full multi-band quantities), the non-local
@@ -224,7 +228,7 @@ def write_to_full_4pt_quantity(obj_full, obj_ineq: LocalFourPoint, sl: slice):
             :return: The full object with this atom's block filled in.
             """
             if obj_full is None:
-                obj_full = deepcopy(obj_ineq)
+                obj_full = obj_ineq.copy()
                 obj_full.mat = np.zeros(
                     (config.sys.n_bands,) * 4 + obj_ineq.current_shape[4:], dtype=obj_ineq.mat.dtype
                 )
@@ -246,7 +250,7 @@ def write_to_full_2pt_quantity(
             :return: The full object with this atom's block filled in.
             """
             if obj_full is None:
-                obj_full = deepcopy(obj_ineq)
+                obj_full = obj_ineq.copy()
                 obj_full.mat = np.zeros(
                     ((1, 1, 1) + (config.sys.n_bands,) * 2 if has_momentum else (config.sys.n_bands,) * 2)
                     + (obj_ineq.current_shape[-1],),
@@ -581,13 +585,27 @@ def autodetect_memory_settings(comm: MPI.Comm) -> None:
         construct_fq_cheap=config.eliashberg.construct_fq_cheap,
     )
 
-    def node_total(distributed: float, single: float, n_ranks: int) -> float:
-        """Memory held on a node with ``n_ranks`` ranks at a branch's peak (see :func:`autodetect_memory_settings`)."""
-        return n_ranks * (baseline + distributed) + single
+    # giwk_full is deduplicated to one copy per node through the whole non-local self-energy section when
+    # config.memory.use_shared_memory_giwk is on (the node root builds it, the others map it read-only; it is re-shared
+    # at the smaller core-box window after the bubble).
+    giwk_shared_credit = (
+        memory_estimator.giwk_baseline_bytes(config.lattice.q_grid.nk_tot, config.sys.n_bands, niv_cut)
+        if config.memory.use_shared_memory_giwk
+        else 0.0
+    )
+    giwk_shared_branches = {"chi0q", "chiq_aux", "sde"}
+
+    def node_total(distributed: float, single: float, n_ranks: int, giwk_shared: bool = False) -> float:
+        """Memory held on a node with ``n_ranks`` ranks at a branch's peak (see :func:`autodetect_memory_settings`).
+        When ``giwk_shared``, the replicated ``giwk_full`` is counted once per node instead of once per rank."""
+        total = n_ranks * (baseline + distributed) + single
+        if giwk_shared:
+            total -= (n_ranks - 1) * giwk_shared_credit
+        return total
 
-    def fits_everywhere(distributed: float, single: float) -> bool:
+    def fits_everywhere(distributed: float, single: float, giwk_shared: bool = False) -> bool:
         """Whether a transient (per-rank ``distributed`` + one-off ``single``) fits the 90% budget on every node."""
-        return all(node_total(distributed, single, r) <= avail * 0.9 for r, avail in nodes.values())
+        return all(node_total(distributed, single, r, giwk_shared) <= avail * 0.9 for r, avail in nodes.values())
 
     flag_to_key = {
         "save_memory_for_chi0q": "chi0q",
@@ -613,17 +631,19 @@ def fits_everywhere(distributed: float, single: float) -> bool:
             continue
         bp = peaks[key]
         label = key_to_label[key]
-        if not fits_everywhere(bp.on_distributed, bp.on_single):
-            worst = max(node_total(bp.on_distributed, bp.on_single, r) for r, _ in nodes.values())
+        # giwk is deduplicated per node across the whole SDE section, so credit those branches (not the eliashberg ones).
+        giwk_shared = key in giwk_shared_branches and config.memory.use_shared_memory_giwk
+        if not fits_everywhere(bp.on_distributed, bp.on_single, giwk_shared):
+            worst = max(node_total(bp.on_distributed, bp.on_single, r, giwk_shared) for r, _ in nodes.values())
             raise MemoryError(
                 f"Even the most memory-lean path for '{label}' needs {worst / 1024**3:.3f} GB on a node, which "
                 f"exceeds 90% of that node's available memory. Use more nodes, fewer ranks per node, a smaller "
                 f"frequency box or k-grid."
             )
-        autodetect_on = not fits_everywhere(bp.off_distributed, bp.off_single)
+        autodetect_on = not fits_everywhere(bp.off_distributed, bp.off_single, giwk_shared)
         final = bool(getattr(config.memory, attr)) or autodetect_on
         setattr(config.memory, attr, final)
-        worst_off = max(node_total(bp.off_distributed, bp.off_single, r) for r, _ in nodes.values())
+        worst_off = max(node_total(bp.off_distributed, bp.off_single, r, giwk_shared) for r, _ in nodes.values())
         logger.info(
             f"{label}: fast-path node total {worst_off / 1024**3:.3f} GB -> "
             f"memory saving {'enabled' if final else 'disabled'}."
@@ -702,4 +722,4 @@ def configure_matplotlib():
 
 
 if __name__ == "__main__":
-    execute_dga_routine()
+    main()

dgamore/__init__.py

@@ -1,6 +1,5 @@
 # SPDX-FileCopyrightText: 2025-2026 Julian Peil <julian.peil@tuwien.ac.at>
 # SPDX-License-Identifier: MIT
 #
-# DGAmore — Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
+# DGAmore - Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
 #           Eliashberg Equation Solver for Strongly Correlated Electron Systems
-

dgamore/brillouin_zone.py

@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: 2025-2026 Julian Peil <julian.peil@tuwien.ac.at>
 # SPDX-License-Identifier: MIT
 #
-# DGAmore — Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
+# DGAmore - Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
 #           Eliashberg Equation Solver for Strongly Correlated Electron Systems
 
 """
@@ -361,7 +361,7 @@ def get_lattice_symmetries_from_string(symmetry_string: str | tuple | list) -> l
 
 
 # Sentinel object returned by get_lattice_symmetries_from_string for "auto".
-# Identity-checked, so it must be a unique singleton — a small dedicated object.
+# Identity-checked, so it must be a unique singleton - a small dedicated object.
 class _AutoSymmetriesSentinel:
     """Marker indicating that lattice symmetries are to be detected automatically
     from a Hamiltonian, via KGrid.specify_auto_symmetries()."""

dgamore/bubble_gen.py

@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: 2025-2026 Julian Peil <julian.peil@tuwien.ac.at>
 # SPDX-License-Identifier: MIT
 #
-# DGAmore — Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
+# DGAmore - Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
 #           Eliashberg Equation Solver for Strongly Correlated Electron Systems
 r"""
 Generalized bare susceptibilities (the "bubbles"). :class:`BubbleGenerator` builds the products of two Green's

dgamore/config.py

@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: 2025-2026 Julian Peil <julian.peil@tuwien.ac.at>
 # SPDX-License-Identifier: MIT
 #
-# DGAmore — Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
+# DGAmore - Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
 #           Eliashberg Equation Solver for Strongly Correlated Electron Systems
 """
 Global configuration singleton. This module holds the process-wide mutable state of a DGAmore run as module-level
@@ -278,6 +278,9 @@ class MemoryConfig:
     :ivar bool save_memory_for_sde: Use the q-loop self-energy contraction instead of the FFT one.
     :ivar bool save_memory_for_fq: Use the per-q full-vertex construction in the Eliashberg step.
     :ivar bool save_memory_for_lanczos: Use the frequency-distributed Lanczos solver.
+    :ivar bool use_shared_memory_giwk: Store the replicated full-grid Green's function ``giwk_full`` in one MPI
+        shared-memory window per node (computed only by the node root) instead of one private copy per rank. Enabled
+        by default; disable it if cross-socket (NUMA) reads of the shared buffer outweigh the memory saving.
     """
 
     def __init__(self):
@@ -286,6 +289,7 @@ def __init__(self):
         self.save_memory_for_sde: bool = False
         self.save_memory_for_fq: bool = False
         self.save_memory_for_lanczos: bool = False
+        self.use_shared_memory_giwk: bool = True
 
 
 class AnaContConfig:

dgamore/config_parser.py

@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: 2025-2026 Julian Peil <julian.peil@tuwien.ac.at>
 # SPDX-License-Identifier: MIT
 #
-# DGAmore — Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
+# DGAmore - Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
 #           Eliashberg Equation Solver for Strongly Correlated Electron Systems
 """
 YAML configuration parsing. :class:`ConfigParser` reads the run's YAML config (rank 0), broadcasts it to all MPI
@@ -326,6 +326,7 @@ def _build_memory_config(self, config_file):
         conf.save_memory_for_sde = self._try_parse(section, "save_memory_for_sde", conf.save_memory_for_sde)
         conf.save_memory_for_fq = self._try_parse(section, "save_memory_for_fq", conf.save_memory_for_fq)
         conf.save_memory_for_lanczos = self._try_parse(section, "save_memory_for_lanczos", conf.save_memory_for_lanczos)
+        conf.use_shared_memory_giwk = self._try_parse(section, "use_shared_memory_giwk", conf.use_shared_memory_giwk)
 
         return conf
 

dgamore/dga_config.yaml

@@ -96,4 +96,6 @@ memory:
   # less memory intensive, but it will increase runtime.
   save_memory_for_lanczos: False # If True, the code will perform the Lanczos algorithm with the pairing vertex for
   # the Eliashberg calculation in a more memory efficient way. This will reduce memory usage but increase runtime
-  # due to required MPI communication in each Lanczos step.
+  # due to required MPI communication in each Lanczos step.
+  use_shared_memory_giwk: True # If True (default), the full-grid lattice Green's function is stored once per node in
+  # an MPI shared-memory window (computed only by the node's first rank)

dgamore/dga_io.py

@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: 2025-2026 Julian Peil <julian.peil@tuwien.ac.at>
 # SPDX-License-Identifier: MIT
 #
-# DGAmore — Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
+# DGAmore - Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
 #           Eliashberg Equation Solver for Strongly Correlated Electron Systems
 """
 High-level DMFT input loading and run setup. This module ties the DMFT interface, the global config and the lattice

dgamore/dga_logger.py

@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: 2025-2026 Julian Peil <julian.peil@tuwien.ac.at>
 # SPDX-License-Identifier: MIT
 #
-# DGAmore — Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
+# DGAmore - Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
 #           Eliashberg Equation Solver for Strongly Correlated Electron Systems
 """
 MPI-aware logging. :class:`DgaLogger` timestamps messages and, by default, only emits them on the root rank so

dgamore/dmft_interface.py

@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: 2025-2026 Julian Peil <julian.peil@tuwien.ac.at>
 # SPDX-License-Identifier: MIT
 #
-# DGAmore — Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
+# DGAmore - Multi-Orbital Ladder Dynamical Vertex Approximation (LDGA) &
 #           Eliashberg Equation Solver for Strongly Correlated Electron Systems
 """
 DMFT input interface. :class:`DMFTInterface` is the abstract contract for reading the quantities a DGA run needs