Merge pull request #58 from NVIDIA/release/v0.9.0

release: nvmath-python-0.9.0

Author: markjolah

.gitignore

@@ -28,3 +28,17 @@ Pipfile*
 .pipenv-cache
 
 .mypy_cache
+
+# Cursor Local Agent Settings
+.cursor/local/*
+!.cursor/local/settings.template.md
+.cursor/sandbox.json
+.cursor/plans/
+.cursor/rules/
+!.cursor/rules/global-*
+
+# GPU Bridge temporary files
+internal/cursor/run/
+
+# graphify
+graphify-out/

README.md

@@ -225,8 +225,10 @@ shape = 512 // nranks, 256, 512
 a = nvmath.distributed.allocate_symmetric_memory(shape, cp, dtype=cp.complex128)
 # a is a cupy ndarray and can be operated on using in-place cupy operations.
 with cp.cuda.Device(device_id):
-    a[:] = cp.random.rand(*shape, dtype=cp.float64) + 1j *
+    a[:] = (
         cp.random.rand(*shape, dtype=cp.float64)
+        + 1j * cp.random.rand(*shape, dtype=cp.float64)
+    )
 
 # Forward FFT.
 # In this example, the forward FFT operand is distributed according

builder/utils.py

@@ -20,13 +20,12 @@ def detect_cuda_paths():
     # isolation is on, the build prefix is added to sys.path, but this is the only
     # implementation detail that we rely on.
     # TODO: move to cuda.pathfinder.
-    potential_build_prefixes = (
-        [os.path.join(p, "nvidia/cu13") for p in sys.path]
-        # internal/bindings depends on cuda_bindings cydriver,
-        # which introduces dependency on cudaProfiler.h
-        + [os.path.join(p, "nvidia/cuda_profiler_api") for p in sys.path]
-        + [os.environ.get("CUDA_PATH", os.environ.get("CUDA_HOME", "")), "/usr/local/cuda"]
-    )
+    potential_build_prefixes = [os.path.join(p, "nvidia/cu13") for p in sys.path] + [
+        os.environ.get("CUDA_PATH", os.environ.get("CUDA_HOME", "")),
+        "/usr/local/cuda",
+    ]
+    # Filter out empty paths, e.g. when neither CUDA_PATH nor CUDA_HOME is set
+    potential_build_prefixes = [p for p in potential_build_prefixes if p]
     cuda_paths = []
 
     def check_path(header):
@@ -46,9 +45,6 @@ def check_path(header):
 
     check_path("cuda.h")
     check_path("crt/host_defines.h")
-    # internal/bindings depends on cuda_bindings cydriver,
-    # which introduces dependency on cudaProfiler.h
-    check_path("cudaProfiler.h")
     return cuda_paths
 
 
@@ -64,7 +60,6 @@ def decide_lib_name(ext_name):
         "curand",
         "nvpl",
         "nvshmem",
-        "nccl",
         "mathdx",
         "cudss",
         "cutensor",

docs/Makefile

@@ -17,6 +17,7 @@ clean:
 	rm -rf ${SOURCEDIR}/bindings/generated/
 	rm -rf ${SOURCEDIR}/host-apis/**/generated
 	rm -rf ${SOURCEDIR}/device-apis/generated
+	rm -rf ${SOURCEDIR}/distributed-apis/generated
 	rm -rf ${SOURCEDIR}/distributed-apis/**/generated
 
 html: Makefile

docs/sphinx/_static/nvmath_override.css

@@ -1,3 +1,43 @@
+/*
+ * sphinx-tabs ships tabs.css for themes that set body[data-theme=...] (e.g. Furo).
+ * PyData / NVIDIA Sphinx theme uses html[data-bs-theme=light|dark|auto] instead, so
+ * those dark rules never match and tab panels / the active tab stay white in dark mode.
+ * tabs.css may override generic rules, so we use .sphinx-tabs ... selectors for
+ * higher specificity and Bootstrap surface tokens that track the color mode switcher.
+ */
+/* Tab row divider: keep separation line aligned with theme border color. */
+.sphinx-tabs [role="tablist"] {
+  border-bottom: 1px solid var(--pst-color-border, #a0b3bf);
+}
+
+/* Inactive tab labels: use theme link color and transparent background. */
+.sphinx-tabs .sphinx-tabs-tab {
+  color: var(--pst-color-link, #1d5c87);
+  background-color: transparent;
+}
+
+/* Active tab: connect visually to panel and inherit theme surface/text colors. */
+.sphinx-tabs .sphinx-tabs-tab[aria-selected="true"] {
+  font-weight: 700;
+  border: 1px solid var(--pst-color-border, #a0b3bf);
+  /* Match panel background to hide the seam between tab and panel. */
+  border-bottom: 1px solid var(--pst-color-surface, #fff);
+  /* Preserve sphinx-tabs geometry so the active tab overlaps panel border cleanly. */
+  margin: -1px;
+  background-color: var(--pst-color-surface, #fff);
+  color: var(--pst-color-text-base, #212529);
+}
+
+/* Tab panel container: theme-aware border/surface/text with original rounded geometry. */
+.sphinx-tabs .sphinx-tabs-panel {
+  border: 1px solid var(--pst-color-border, #a0b3bf);
+  margin: 0 -1px -1px -1px;
+  border-radius: 0 0 5px 5px;
+  border-top: 0;
+  background-color: var(--pst-color-surface, #fff);
+  color: var(--pst-color-text-base, #212529);
+}
+
 /* Make sidebar wider to accommodate our long module names. */
 .bd-sidebar-primary {
     flex-basis: 25%;  /* default is 20% */

docs/sphinx/_static/switcher.json

@@ -1,42 +1,46 @@
 [
   {
     "version": "latest",
-    "url": "https://docs.nvidia.com/cuda/nvmath-python/latest"
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/latest/"
+  },
+  {
+    "version": "0.9.0",
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.9.0/"
   },
   {
     "version": "0.8.0",
-    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.8.0"
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.8.0/"
   },
   {
     "version": "0.7.0",
-    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.7.0"
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.7.0/"
   },
   {
     "version": "0.6.0",
-    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.6.0"
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.6.0/"
   },
   {
     "version": "0.5.0",
-    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.5.0"
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.5.0/"
   },
   {
     "version": "0.4.0",
-    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.4.0"
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.4.0/"
   },
   {
     "version": "0.3.0",
-    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.3.0"
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.3.0/"
   },
   {
     "version": "0.2.1",
-    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.2.1"
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.2.1/"
   },
   {
     "version": "0.2.0",
-    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.2.0"
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.2.0/"
   },
   {
     "version": "0.1.0",
-    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.1.0"
+    "url": "https://docs.nvidia.com/cuda/nvmath-python/0.1.0/"
   }
 ]

docs/sphinx/_templates/autosummary/class.rst

@@ -1,26 +1,35 @@
+{# This template replaces the default autosummary Sphinx template
+
+   - The default template does not show all methods and attributes on one page.
+   - Attributes are listed before methods
+   - __init__ is skipped because the function signature is already shown
+#}
 {{ fullname | escape | underline}}
 
 .. currentmodule:: {{ module }}
 
 .. autoclass:: {{ objname }}
 
-   {% block methods %}
-   {% if methods %}
-   .. rubric:: {{ _('Methods') }}
+   {% block attributes %}
+   {% if attributes %}
+   .. rubric:: {{ _('Attributes') }}
 
-   {% for item in methods %}
-   .. automethod:: {{ item }}
+   {% for item in attributes %}
+   .. autoattribute:: {{ item }}
    {%- endfor %}
 
    {% endif %}
    {% endblock %}
 
-   {% block attributes %}
-   {% if attributes %}
-   .. rubric:: {{ _('Attributes') }}
+   {% block methods %}
+   {% if methods %}
+   .. rubric:: {{ _('Methods') }}
 
-   {% for item in attributes %}
-   .. autoattribute:: {{ item }}
+   {% for item in methods %}
+   {% if item != '__init__' %}
+   .. automethod:: {{ item }}
+   {% endif %}
    {%- endfor %}
+
    {% endif %}
    {% endblock %}

docs/sphinx/_templates/autosummary/dataclass.rst

@@ -5,5 +5,4 @@
 .. autoclass:: {{ objname }}
 
    {% block methods %}
-   .. automethod:: __init__
    {% endblock %}

docs/sphinx/_templates/autosummary/torchust.rst

@@ -0,0 +1,12 @@
+{# Autosummary template for TorchUST.
+
+   - Only from_torch is documented because it is the primary public entry point;
+     other inherited members are internal or covered elsewhere.
+#}
+{{ fullname | escape | underline}}
+
+.. currentmodule:: {{ module }}
+
+.. autoclass:: {{ objname }}
+   :no-inherited-members:
+   :members: from_torch

docs/sphinx/api_design.rst

@@ -0,0 +1,8 @@
+API Design and Usage Patterns
+*****************************
+
+.. toctree::
+   :maxdepth: 2
+
+   Stateful APIs <stateful_apis_guide.rst>
+   Stream Semantics <stream_semantics.rst>