docs: contributing guide, PR template, CODEOWNERS; build trees in build/<preset>

.gitattributes

@@ -0,0 +1,12 @@
+* text=auto
+*.png binary
+*.jpg binary
+*.npy binary
+*.npz binary
+*.safetensors binary
+*.gguf binary
+*.bin binary
+third_party/** linguist-vendored
+tests/data/** linguist-generated
+tests/golden/** linguist-generated
+tests/reference_images/** linguist-generated

.github/CODEOWNERS

@@ -0,0 +1,2 @@
+# Review requests for pull requests. The last matching line wins.
+* @xycld

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,11 @@
+## What changes
+
+<!-- One or two sentences. Link the issue if there is one. -->
+
+## Tests
+
+<!-- CI builds every package and runs the CPU tests. Name what else ran: GPU tests, which GPU, model files. -->
+
+## Speed and memory
+
+<!-- For a change on a hot path: s/step and peak VRAM before and after, same GPU, model and settings. Otherwise "none". -->

.github/workflows/build.yml

@@ -249,29 +249,29 @@ jobs:
           ctest --preset "$preset" -L '^cpu$' --parallel 4
 
       - name: Install
-        run: cmake --install "build-$preset" --prefix "build-$preset/install"
+        run: cmake --install "build/$preset" --prefix "build/$preset/install"
 
       - name: Install the system-CUDA variant
         # Everything but the `cuda` component (cmake/modules/packaging.cmake): no cuBLASLt, no CUDA EULA.
-        run: cmake --install "build-$preset" --component Unspecified --prefix "build-$preset/install-system-cuda"
+        run: cmake --install "build/$preset" --component Unspecified --prefix "build/$preset/install-system-cuda"
 
       - name: AC1 / exports / AC17 / find_package consumer
         run: |
           cuda="${{ matrix.cuda }}"
-          python .github/scripts/check_package.py --prefix "build-$preset/install" --cuda-major "${cuda%%.*}" \
-            --consumer examples/consumer-cmake --consumer-build "build-$preset/consumer" --run-consumer
+          python .github/scripts/check_package.py --prefix "build/$preset/install" --cuda-major "${cuda%%.*}" \
+            --consumer examples/consumer-cmake --consumer-build "build/$preset/consumer" --run-consumer
 
       - name: AC1 / exports / AC17 / no CUDA libraries (system-CUDA variant)
-        run: python .github/scripts/check_package.py --prefix "build-$preset/install-system-cuda" --system-cuda
+        run: python .github/scripts/check_package.py --prefix "build/$preset/install-system-cuda" --system-cuda
 
       - name: Package
-        working-directory: build-${{ env.preset }}
+        working-directory: build/${{ env.preset }}
         run: cpack --config CPackConfig.cmake
 
       - name: Package the system-CUDA variant
         # The CPack archives' layout (one top-level directory named like the archive) and checksum files, from the
         # system-CUDA install: mossinfer-<version>-<os>-cu<N>-<arch>-system-cuda.{zip,tar.gz}.
-        working-directory: build-${{ env.preset }}
+        working-directory: build/${{ env.preset }}
         run: |
           full=$(ls mossinfer-*.zip | grep -v -- '-system-cuda\.zip$')
           [ "$(echo "$full" | wc -l)" -eq 1 ] || { echo "expected one CPack zip, got: $full"; exit 1; }
@@ -290,7 +290,7 @@ jobs:
 
       - name: AC17 size of the archive
         run: |
-          for archive in build-"$preset"/mossinfer-*.zip build-"$preset"/mossinfer-*.tar.gz; do
+          for archive in build/"$preset"/mossinfer-*.zip build/"$preset"/mossinfer-*.tar.gz; do
             [ -f "$archive" ] || continue
             case "$archive" in
               *-system-cuda.*) python .github/scripts/check_package.py --archive "$archive" --system-cuda ;;
@@ -304,9 +304,9 @@ jobs:
         with:
           name: mossinfer-${{ matrix.platform }}-${{ matrix.flavour }}
           path: |
-            build-${{ env.preset }}/mossinfer-*.zip
-            build-${{ env.preset }}/mossinfer-*.tar.gz
-            build-${{ env.preset }}/mossinfer-*.sha256
+            build/${{ env.preset }}/mossinfer-*.zip
+            build/${{ env.preset }}/mossinfer-*.tar.gz
+            build/${{ env.preset }}/mossinfer-*.sha256
           if-no-files-found: error
           retention-days: ${{ inputs.retention-days }}
           compression-level: 0

.mailmap

@@ -0,0 +1 @@
+xycld <xycloud.cn@gmail.com> xy_cloud <85053859+xycld@users.noreply.github.com>

CMakePresets.json

@@ -6,8 +6,8 @@
       "name": "base",
       "hidden": true,
       "generator": "Ninja",
-      "binaryDir": "${sourceDir}/build-${presetName}",
-      "installDir": "${sourceDir}/build-${presetName}/install",
+      "binaryDir": "${sourceDir}/build/${presetName}",
+      "installDir": "${sourceDir}/build/${presetName}/install",
       "cacheVariables": { "CMAKE_EXPORT_COMPILE_COMMANDS": "ON" }
     },
     {
@@ -97,7 +97,7 @@
       "name": "windows-msvc-cuda13-sm75",
       "displayName": "Windows release, CUDA 13.x with Turing (sm_75)",
       "inherits": "windows-msvc-cuda13",
-      "binaryDir": "${sourceDir}/build-${presetName}",
+      "binaryDir": "${sourceDir}/build/${presetName}",
       "cacheVariables": { "MOSSINFER_SM75": "ON" }
     },
     {

CONTRIBUTING.md

@@ -0,0 +1,51 @@
+# Contributing
+
+## Build and test
+
+CUDA 12.x or 13.x, CMake ≥ 3.28 and Ninja; on Linux GCC ≥ 12 and GoogleTest, on Windows a Visual Studio 2022 x64
+developer prompt.
+
+```sh
+cmake --preset dev && cmake --build --preset dev
+CUDA_VISIBLE_DEVICES="" ctest --preset dev -L '^cpu$'     # what CI runs
+ctest --preset dev                                         # adds the GPU tests on your GPU
+```
+
+The `dev` preset builds for compute capability 8.9; pass `-DCMAKE_CUDA_ARCHITECTURES=<cc>` for another GPU. Build
+trees go to `build/<preset>`. Tests that need model files skip when the files are absent.
+
+CI builds every release package (Linux and Windows, CUDA 12 and 13, the `-sm75` and `-sm70` builds, Vulkan and HIP)
+and runs the CPU tests and the package checks. It has no GPU: a change to a kernel, the planner or a pipeline path
+needs a GPU run, named in the pull request.
+
+## Branches
+
+Branch from `main` and name the branch for the kind of change: `feat/`, `fix/`, `perf/`, `refactor/`, `docs/`, `ci/`,
+`test/`, and `release/X.Y.Z` for a version bump. Open a draft pull request to run CI on work in progress.
+
+## Commits
+
+Subject line: `area: what changes`, in the imperative, without a final period. Areas: `cuda`, `kernels`, `pipeline`,
+`planner`, `weights`, `quantize`, `capi`, `server`, `cli`, `sm75`, `sm70`, `vulkan`, `hip`, `build`, `ci`, `release`,
+`docs`, `tests`.
+
+```
+vulkan: count the DiT's weights in --vram-budget with every block streamed
+capi: report the preload's plan before its first phase
+```
+
+The body says why, and what the change costs in speed or memory if anything.
+
+## Pull requests
+
+`main` takes changes through pull requests only. Small changes are squash-merged with the pull request's title as the
+commit subject; a branch whose commits each build and tell a story (a new backend, a series of kernels) is merged with
+a merge commit. A change on a hot path reports s/step and peak VRAM before and after, on the same GPU, model and
+settings: no feature may cost speed.
+
+## Releases
+
+Maintainers release from `main`. A `release/X.Y.Z` pull request bumps `MI_VERSION_*` in
+`include/mossinfer/mossinfer.h` and the version strings in the docs. After it merges and CI passes on `main`, a signed
+`vX.Y.Z` tag on the merge commit starts `.github/workflows/release.yml`, which builds, checks and publishes the
+packages. `vX.Y.Z-rcN` tags publish pre-releases. Release tags cannot be moved or deleted.

docs/integration.md

@@ -203,14 +203,14 @@ On top of that, the weights are read through the operating system's page cache.
 cmake --preset linux-cuda13
 cmake --build --preset linux-cuda13 -j8
 CUDA_VISIBLE_DEVICES="" ctest --preset linux-cuda13 -L '^cpu$'
-cmake --install build-linux-cuda13                # -> build-linux-cuda13/install
-cmake --install build-linux-cuda13 --component Unspecified --prefix build-linux-cuda13/install-system-cuda
-(cd build-linux-cuda13 && cpack --config CPackConfig.cmake)   # -> mossinfer-2.2.0-linux-cu13-x86_64.{tar.gz,zip}
+cmake --install build/linux-cuda13                # -> build/linux-cuda13/install
+cmake --install build/linux-cuda13 --component Unspecified --prefix build/linux-cuda13/install-system-cuda
+(cd build/linux-cuda13 && cpack --config CPackConfig.cmake)   # -> mossinfer-2.2.0-linux-cu13-x86_64.{tar.gz,zip}
 ```
 
 - **glibc:** the archive needs the glibc of the machine that built it, or a newer one. Build on the oldest distribution you want to support. CI uses Ubuntu 22.04, which gives glibc 2.35. A build from a rolling distribution such as Arch will not start on Ubuntu LTS.
 - **Linux CUDA 12 with Volta:** use the `linux-cuda12-sm70` configure/build/test/package presets. This produces a `-sm70` package; CUDA 13 cannot build or run the Volta backend. See [sm70.md](sm70.md#build).
-- **Windows CUDA 13 with Turing:** from a Visual Studio x64 developer prompt, run `cmake --preset windows-msvc-cuda13-sm75`, `cmake --build --preset windows-msvc-cuda13-sm75`, `ctest --preset windows-msvc-cuda13-sm75`, then `cpack --config build-windows-msvc-cuda13-sm75/CPackConfig.cmake -B build-windows-msvc-cuda13-sm75`. This creates the full `-sm75.zip` package; the ordinary presets remain Ampere-and-newer. Any preset takes `-DMOSSINFER_SM75=ON`; what the build adds is in [sm75.md](sm75.md#build).
+- **Windows CUDA 13 with Turing:** from a Visual Studio x64 developer prompt, run `cmake --preset windows-msvc-cuda13-sm75`, `cmake --build --preset windows-msvc-cuda13-sm75`, `ctest --preset windows-msvc-cuda13-sm75`, then `cpack --config build/windows-msvc-cuda13-sm75/CPackConfig.cmake -B build/windows-msvc-cuda13-sm75`. This creates the full `-sm75.zip` package; the ordinary presets remain Ampere-and-newer. Any preset takes `-DMOSSINFER_SM75=ON`; what the build adds is in [sm75.md](sm75.md#build).
 - **Other presets:** `linux-cuda12`, `windows-msvc-cuda13` and `windows-msvc-cuda12`. The Windows presets run from a Visual Studio 2022 x64 developer prompt.
 
 ### Package contents

docs/sm70.md

@@ -15,10 +15,10 @@ machine whose supported devices are all Volta, the probe recommends `cu12` even
 cmake --preset linux-cuda12-sm70 -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.9/bin/nvcc
 cmake --build --preset linux-cuda12-sm70
 CUDA_VISIBLE_DEVICES="" ctest --preset linux-cuda12-sm70 -L '^cpu$'
-build-linux-cuda12-sm70/sm70_check all
-build-linux-cuda12-sm70/sm70_check big
-build-linux-cuda12-sm70/mossinfer_tests --gtest_filter='Sm70Build.*:GpuSm70.*:GpuSm75Convert.*'
-cpack --config build-linux-cuda12-sm70/CPackConfig.cmake -B build-linux-cuda12-sm70
+build/linux-cuda12-sm70/sm70_check all
+build/linux-cuda12-sm70/sm70_check big
+build/linux-cuda12-sm70/mossinfer_tests --gtest_filter='Sm70Build.*:GpuSm70.*:GpuSm75Convert.*'
+cpack --config build/linux-cuda12-sm70/CPackConfig.cmake -B build/linux-cuda12-sm70
 ```
 
 The package name ends in `-sm70`. The build/release matrix includes the Linux CUDA 12 SM70 preset. The CUDA architecture list of the ordinary core must stay sm_80 or newer; do not

docs/sm75.md

@@ -16,8 +16,8 @@ The Windows CUDA 13 preset builds a complete package and names it with a `-sm75`
 cmake --preset windows-msvc-cuda13-sm75
 cmake --build --preset windows-msvc-cuda13-sm75
 ctest --preset windows-msvc-cuda13-sm75 -L '^cpu$'
-cpack --config build-windows-msvc-cuda13-sm75/CPackConfig.cmake -B build-windows-msvc-cuda13-sm75
-# -> build-windows-msvc-cuda13-sm75/mossinfer-2.2.0-windows-cu13-x86_64-sm75.zip
+cpack --config build/windows-msvc-cuda13-sm75/CPackConfig.cmake -B build/windows-msvc-cuda13-sm75
+# -> build/windows-msvc-cuda13-sm75/mossinfer-2.2.0-windows-cu13-x86_64-sm75.zip
 ```
 
 Any other preset takes the option too (`-DMOSSINFER_SM75=ON`), e.g. a Linux build from a custom preset; the ordinary

tests/server/test_install_tree.py

@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 """Install-tree test of the cuBLASLt search (src/backend/cuda_libs.h). Standard library only; needs no GPU.
 
-    python3 tests/server/test_install_tree.py --build build-dev --work build-dev/install_tree_test \\
+    python3 tests/server/test_install_tree.py --build build/dev --work build/dev/install_tree_test \\
         [--check-package .github/scripts/check_package.py]
 
 Installs the build twice: the full tree (every component, cuBLASLt bundled) and the system-CUDA variant