Reduce installed Open3D wheel size

.github/agents/open3d-issue-fixer.agent.md

@@ -0,0 +1,78 @@
+---
+name: "Open3D Issue Fixer"
+description: "Use when: investigating and fixing an Open3D GitHub issue from an issue number, URL, or description; validate scope, correctness, reproducibility, complexity, and expected benefit; implement and test a focused fix; then commit and push it for CI, either on the current branch or a dedicated PR branch."
+argument-hint: "Issue number, URL, or description; optionally specify current branch or new PR branch"
+tools: [read, search, edit, execute, todo, agent, web, mcp_github/*]
+agents: ["Universal Janitor"]
+user-invocable: true
+disable-model-invocation: false
+---
+Investigate an Open3D issue, validate it, implement a justified fix, and push a
+tested commit for CI. Follow `AGENTS.md` and `AGENTS.local.md`; the repository
+and current worktree are authoritative.
+
+## Goal Lock
+
+Before editing, record the issue, user-visible goal, acceptance requirements,
+smallest relevant test, risks, and implementation status in a Markdown progress
+file as required by `AGENTS.md`.
+
+## Workflow
+
+1. **Identify.** Accept an issue number, URL, or description. For a description,
+   search open and closed issues for duplicates. Read the full issue, comments,
+   labels, state, linked PRs, and relevant history. Confirm it belongs to this
+   repository before editing.
+
+2. **Validate.** Check scope and correctness against supported behavior, docs,
+   APIs, source, tests, configurations, and issue discussion. Capture the
+   baseline with the smallest deterministic reproducer. Assess affected users,
+   severity, compatibility, implementation and maintenance cost, cross-layer
+   work, and CI/platform cost. Report one evidence-based verdict: `accept`,
+   `needs clarification`, `duplicate`, `cannot reproduce`, `expected behavior`,
+   or `out of scope`. Do not implement unless the verdict supports it or the
+   user explicitly requests further investigation.
+
+3. **Discriminate.** Follow the debugging process in `AGENTS.md`: enumerate
+   root-cause hypotheses, split them into easy and hard to fix, try the easy
+   candidate fixes directly, and instrument before attempting a hard one. Keep
+   branch, edit, test, and git-write decisions in this agent when delegating
+   research.
+
+4. **Choose the branch.** Inspect the branch, active PR, remotes, and worktree.
+   Continue when the issue belongs to the active PR or the user selected the
+   branch. Otherwise ask whether to continue or create a descriptive issue
+   branch; do not choose implicitly. Ask before switching if unrelated changes
+   prevent it.
+
+5. **Fix and verify.** Follow `AGENTS.md` for implementation, coverage, docs,
+   supported backends, style, and validation. Ensure the reproducer passes.
+
+6. **Deliver.** Commit only the fix, exclude the progress file, and end the
+   commit subject with the issue number: e.g. (#9999). Include the verdict in a
+   new PR. For an existing PR, report the pushed commit and CI target instead of
+   opening a duplicate.
+
+## Stop Conditions
+
+Stop without committing a speculative fix when:
+
+- the issue is out of scope, invalid, a duplicate, or expected behavior;
+- the report cannot be reproduced after reasonable targeted investigation;
+- benefit does not justify implementation or maintenance cost;
+- acceptance criteria require a product or API decision;
+- required credentials, permissions, hardware, or a clean branch transition are
+  unavailable;
+- focused validation fails for reasons not caused by the proposed change.
+
+Do not create, close, label, or comment on issues, and do not merge a PR, unless
+the user explicitly requests that GitHub mutation.
+
+## Final Report
+
+Summarize:
+
+- issue identity, verdict, reproduction, and effort-versus-benefit assessment;
+- root cause and fix, with file links;
+- tests and style checks, including unavailable configurations;
+- branch, commit, push result, PR target, remaining risks, and CI checks.

.github/prompts/fix-pr.prompt.md

@@ -0,0 +1,54 @@
+---
+name: Fix PR CI and Review Comments
+description: "Use when: triaging failing GitHub Actions checks and reviewer feedback on a pull request or branch; separate actionable failures from infrastructure flakes, judge each review comment on user benefit versus maintenance and complexity cost, fix and validate locally, then commit and push."
+argument-hint: "Optional PR number, URL, or branch; otherwise use the active pull request / current branch"
+agent: "agent"
+tools: [read, search, edit, execute, todo, agent, web, mcp_github/*]
+---
+Triage and fix CI failures and review feedback for a pull request. Follow
+`AGENTS.md` and `AGENTS.local.md`; this prompt only adds PR-specific triage
+rules.
+
+## Workflow
+
+1. **Identify.** Resolve the PR from the argument, else the active PR / current
+   branch. Stop and ask if there is no PR, several candidates, a detached HEAD,
+   or a fork head you cannot push to.
+
+2. **Collect.** Read the checks for the PR head commit and the logs of each
+   failing job. Flag results from an older commit as stale, and unpushed local
+   commits as not yet covered by CI. Read review comments and requested changes,
+   skipping outdated or resolved threads.
+
+3. **Classify each failure** as actionable (code, test, build, style, docs, or
+   config defect), infrastructure (download, runner, timeout, quota, cancelled),
+   or flaky/unrelated (also fails on base, or untouched by the diff). Only fix
+   the first; report the rest without claiming they were fixed. Re-run a check
+   only when a transient cause is likely and re-running is cheap.
+
+4. **Classify each review comment** as valid or non-actionable, with a one-line
+   reason weighing user benefit against maintenance and complexity cost.
+   Non-actionable: subjective, obsolete, unsupported by the code, or outside the
+   PR's scope. Ask instead of guessing when a comment implies an API, product,
+   or scope decision.
+
+5. **Fix and validate.** Reproduce locally first where possible, and follow the
+   debugging process in `AGENTS.md`: try the easy candidate fixes directly, and
+   instrument before attempting a hard or complex one. Stay inside the PR's
+   existing scope; do not opportunistically refactor. Name the configurations
+   (CUDA, SYCL, macOS, Windows) you could not verify locally.
+
+6. **Iterate** until every actionable item is fixed or a concrete blocker
+   remains, then commit and push.
+
+Do not merge or close the PR, and do not post comments, reviews, or thread
+resolutions unless the user explicitly asks.
+
+## Final Report
+
+- actionable CI failures and their root causes;
+- each review comment, marked valid or non-actionable, with the reason;
+- fixes made, with file links;
+- validation commands actually run, their results, and unverified platforms;
+- ignored infrastructure/flaky failures and remaining blockers;
+- commit hash, push result, and the CI run to watch.

.github/workflows/macos.yml

@@ -183,9 +183,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           repository: isl-org/Open3D-ML
-          # TODO(ss/sycl-mlops): switch back to main once the torch 2.13
-          # requirements-torch.txt bump on ss/xpu is merged upstream.
-          ref: ss/xpu
           path: ${{ env.OPEN3D_ML_ROOT }}
 
       - name: Download build artifact
@@ -299,9 +296,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           repository: isl-org/Open3D-ML
-          # TODO(ss/sycl-mlops): switch back to main once the torch 2.13
-          # requirements-torch.txt bump on ss/xpu is merged upstream.
-          ref: ss/xpu
           path: ${{ env.OPEN3D_ML_ROOT }}
 
       - name: Download wheels

.github/workflows/ubuntu-wheel.yml

@@ -243,9 +243,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           repository: isl-org/Open3D-ML
-          # TODO(ss/sycl-mlops): switch back to main once the torch 2.13
-          # requirements-torch.txt bump on ss/xpu is merged upstream.
-          ref: ss/xpu
           path: ${{ env.OPEN3D_ML_ROOT }}
       - name: Download wheels
         uses: actions/download-artifact@v4

.github/workflows/ubuntu.yml

@@ -160,14 +160,16 @@ jobs:
           name: open3d-wheel-linux-x86_64
 
       - name: Checkout Open3D-ML
-        # TODO(ss/sycl-mlops): switch back to main once the torch 2.13
-        # requirements-torch.txt bump on ss/xpu is merged upstream.
-        run: git clone --depth 1 https://github.com/isl-org/Open3D-ML.git -b ss/xpu /tmp/Open3D-ML
+        uses: actions/checkout@v4
+        with:
+          repository: isl-org/Open3D-ML
+          # actions/checkout only accepts paths inside GITHUB_WORKSPACE.
+          path: open3d_ml
 
       - name: Install doc dependencies
         run: |
           source util/ci_utils.sh
-          install_docs_dependencies /tmp/Open3D-ML
+          install_docs_dependencies "${GITHUB_WORKSPACE}/open3d_ml"
 
       - name: Install Open3D wheel
         run: python -m pip install open3d*.whl

.github/workflows/windows.yml

@@ -119,9 +119,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           repository: isl-org/Open3D-ML
-          # TODO(ss/sycl-mlops): switch back to main once the torch 2.13
-          # requirements-torch.txt bump on ss/xpu is merged upstream.
-          ref: ss/xpu
           path: open3d_ml
 
       - name: Install PyTorch requirements
@@ -414,6 +411,8 @@ jobs:
       BUILD_CUDA_MODULE: ${{ matrix.device == 'cuda' && 'ON' || 'OFF' }}
       BUILD_SYCL_MODULE: ${{ matrix.device == 'xpu' && 'ON' || 'OFF' }}
       BUILD_PYTORCH_OPS: ${{ matrix.device == 'cuda' && 'OFF' || 'ON' }}
+      # Bundles the Open3D-ML models and pipelines when ML ops are built.
+      OPEN3D_ML_ROOT: ${{ github.workspace }}/open3d_ml
 
     steps:
 
@@ -437,9 +436,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           repository: isl-org/Open3D-ML
-          # TODO(ss/sycl-mlops): switch back to main once the torch 2.13
-          # requirements-torch.txt bump on ss/xpu is merged upstream.
-          ref: ss/xpu
           path: open3d_ml
 
       - name: Set up Python
@@ -490,6 +486,16 @@ jobs:
           python -m pip install -U -r python/requirements.txt `
             -r python/requirements_build.txt `
             -r python/requirements_jupyter_build.txt
+          if ('${{ env.BUILD_CUDA_MODULE }}' -eq 'ON') {
+            python -m pip install -U -r python/requirements_win_cuda.txt
+          }
+          if ('${{ env.BUILD_SYCL_MODULE }}' -eq 'ON') {
+            python -m pip install -U -r python/requirements_sycl.txt
+          }
+          if (('${{ env.BUILD_TENSORFLOW_OPS }}' -eq 'ON') -or
+              ('${{ env.BUILD_PYTORCH_OPS }}' -eq 'ON')) {
+            python -m pip install -U -r open3d_ml/requirements.txt
+          }
           if ('${{ env.BUILD_PYTORCH_OPS }}' -eq 'ON') {
             if ('${{ matrix.device }}' -eq 'xpu') {
               python -m pip install -U -r open3d_ml/requirements-torch-xpu.txt
@@ -601,7 +607,7 @@ jobs:
           if-no-files-found: error
 
       - name: Update devel release with wheel
-        if: ${{ env.SKIPPED != 'true' && github.ref == 'refs/heads/main' }}
+        if: ${{ env.SKIPPED != 'true' && github.ref == 'refs/heads/main' && matrix.device != 'cuda' }}
         env:
           GH_TOKEN: ${{ github.token }}
         shell: bash
@@ -642,9 +648,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           repository: isl-org/Open3D-ML
-          # TODO(ss/sycl-mlops): switch back to main once the torch 2.13
-          # requirements-torch.txt bump on ss/xpu is merged upstream.
-          ref: ss/xpu
           path: open3d_ml
 
       - name: Download wheels
@@ -683,6 +686,11 @@ jobs:
           }
           echo "Installing Open3D wheel $PIP_PKG_NAME in virtual environment..."
           python -m pip install $PIP_PKG_NAME
+          # Wheels bundling Open3D-ML declare its Python dependencies as the
+          # `ml` extra, needed by the models, pipelines and datasets.
+          $bundlesML = python -c "import open3d; print(open3d._build_config['BUNDLE_OPEN3D_ML'])" |
+            Select-Object -Last 1
+          if ($bundlesML.Trim() -eq 'True') { python -m pip install "${PIP_PKG_NAME}[ml]" }
           python -c "import open3d; print('Imported:', open3d)"
           python -c "import open3d; print('CUDA enabled: ', open3d.core.cuda.is_available())"
           python -c "import open3d; print('SYCL enabled: ', open3d.core.sycl.is_available() if hasattr(open3d.core, 'sycl') else False)"

3rdparty/README_SYCL.md

@@ -111,6 +111,11 @@ with `oneapi-ci <https://github.com/oneapi-src/oneapi-ci/tree/0804a4c9281440d8a9
 User:
 - https://pypi.org/user/IntelAutomationEngineering/
 
+Linux import without sourcing oneAPI: `pip install -r python/requirements_sycl.txt`
+(transitive dpcpp-cpp-rt / onemkl-sycl-* DSOs land in `<venv>/lib`). With
+`BUILD_SYCL_MODULE`, extension RPATH (and wheel patchelf when available) points
+at `<Python prefix>/lib`.
+
 Libraries:
 - dpcpp-cpp-rt     https://pypi.org/project/dpcpp-cpp-rt/#history
 - mkl              https://pypi.org/project/mkl/#history

3rdparty/cmake/FindPytorch.cmake

@@ -18,6 +18,8 @@
 macro(translate_arch_string input output)
     if("${input}" MATCHES "[0-9]+-real")
         string(REGEX REPLACE "([1-9])([0-9])-real" "\\1.\\2" version "${input}")
+    elseif("${input}" MATCHES "[0-9]+-virtual")
+        string(REGEX REPLACE "([1-9])([0-9])-virtual" "\\1.\\2+PTX" version "${input}")
     elseif("${input}" MATCHES "([0-9]+)")
         string(REGEX REPLACE "([1-9])([0-9])" "\\1.\\2+PTX" version "${input}")
     elseif("${input}" STREQUAL "native")

3rdparty/find_dependencies.cmake

@@ -1542,9 +1542,11 @@ if(BUILD_GUI)
             # We first search for these paths, and then search CMake's default
             # search path. LLVM version must be >= 7 to compile Filament.
             if (NOT CLANG_LIBDIR)
-                message(STATUS "Searching /usr/lib/llvm-[7..19]/lib/ for libc++ and libc++abi")
-                foreach(llvm_ver RANGE 7 19)
+                message(STATUS "Searching /usr/lib/llvm-[7..20]/lib/ for libc++ and libc++abi")
+                foreach(llvm_ver RANGE 7 20)
                     set(llvm_lib_dir "/usr/lib/llvm-${llvm_ver}/lib")
+                    unset(CPP_LIBRARY CACHE)
+                    unset(CPPABI_LIBRARY CACHE)
                     find_library(CPP_LIBRARY    c++ PATHS ${llvm_lib_dir} NO_DEFAULT_PATH)
                     find_library(CPPABI_LIBRARY c++abi PATHS ${llvm_lib_dir} NO_DEFAULT_PATH)
                     if (CPP_LIBRARY AND CPPABI_LIBRARY)
@@ -1561,6 +1563,7 @@ if(BUILD_GUI)
             if (NOT CLANG_LIBDIR)
                 message(STATUS "Clang C++ libraries not found. Searching other paths...")
                 find_library(CPPABI_LIBRARY c++abi PATH_SUFFIXES
+                             llvm-20/lib
                              llvm-19/lib
                              llvm-18/lib
                              llvm-17/lib
@@ -1592,8 +1595,8 @@ if(BUILD_GUI)
 
             # Ensure that libstdc++ gets linked first.
             target_link_libraries(3rdparty_filament INTERFACE -lstdc++
-                                  ${CPP_LIBRARY}.1 ${CPPABI_LIBRARY})
-            message(STATUS "Filament C++ libraries: ${CPP_LIBRARY}.1 ${CPPABI_LIBRARY}")
+                                  ${CPP_LIBRARY} ${CPPABI_LIBRARY})
+            message(STATUS "Filament C++ libraries: ${CPP_LIBRARY} ${CPPABI_LIBRARY}")
             if (LIBCPP_VERSION GREATER 11)
                 message(WARNING "libc++ (LLVM) version ${LIBCPP_VERSION} > 11 includes libunwind that "
                 "interferes with the system libunwind.so.8 and may crash Python code when exceptions "