Support for Apple Silicon (Metal/MPS) - #357
Open
fnachon wants to merge 8 commits into
Open
Conversation
Apple Silicon-compatible fork created with Claude. Use Python 3.13 environment. This fork was created for use with a mps fork of Proteina-Complexa.
…n crash - Add CMake/ninja/scikit-build artifacts to .gitignore (.cmake/reply/, CMakeFiles/, CMakeCache.txt, build.ninja, .ninja_*, .skbuild-info.json, metal_air/, Makefile, cmake_install.cmake, pack_rotamers_*.pdb) - Untrack 108 previously-committed build artifacts via git rm --cached - Fix _cpp_lib.py: use glob to locate _C.so instead of importlib.util.find_spec (find_spec executes the .so init as a side effect, causing double-registration of TORCH_LIBRARY ops and a C++ abort on subsequent load_library calls); add sys.modules guard as additional safety Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Set KMP_DUPLICATE_LIB_OK=TRUE in execute_process calls that import torch so the duplicate-libomp crash no longer prevents resolving the conda torch cmake prefix path. Also auto-discover pybind11's cmake dir from the active Python environment the same way, replacing the need for a system-installed pybind11. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The KMP_DUPLICATE_LIB_OK=TRUE workaround was missing from the execute_process that fetches the torch version, causing an abort and an empty _TORCH_VERSION list → "list GET given empty list" errors. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Merges 40 commits from uw-ipd/tmol (packer rewrite, ContextManager perf refactor, genbonded score term, coordinate constraints, PyTorch 2.13/C++20 support, CPU-only builds, new wheel matrix) into the MPS fork. Resolves 19 conflicting files plus several silent auto-merge issues found by rebuilding and testing: - CMakeLists.txt: combined upstream's generic CUDA/CPU detection with the MPS backend; fixed a latent bug where cartbonded/genbonded .cuda.cu files were unconditionally listed (breaks CPU-only builds). - Consolidated the standalone compiled_inverse_kin pybind module into the unified compiled_ops TORCH_LIBRARY (matching upstream's refactor); kept the MPS CPU-fallback wrapper in kinematics/compiled/__init__.py. - Added ContextManager& parameters throughout device_operations.mps.impl.hh and the kinematics/pack MPS dispatch copies to match upstream's ContextManager perf refactor (mgpu context reuse). - Added missing MPS instantiations and mps_to_dev() round-trips for the two score terms upstream added without MPS awareness (genbonded, hbond's gen_hbond_bases) - fixes a crash from concatenating a stray CPU-resident tensor with MPS tensors in whole-pose scoring. - Fixed _build_static_tails (scoring_module.py) crashing on MPS by building float64 tails only for devices that support float64. - Fixed a stray merge artifact in impose_rotamers.py referencing an undefined `assignment` variable left over from the old function signature. - Dropped the test_custom_op CMake target after upstream removed its source files. Verified: full CPU+MPS rebuild succeeds; all 23 dedicated MPS tests pass; broad CPU sweep across kinematics/pack/score/optimization/relax passes 515/544 non-deselected tests, with remaining failures confirmed pre-existing (PyTorch 2.11 gradcheck check_undefined_grad_support segfault, reproduced on pre-merge commit c70bc78) or requiring a CUDA build (upstream test extensions gated behind TMOL_HAS_CUDA). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Collaborator
|
This is not merge-ready against the current build matrix. Concrete blockers found in the audit: it accidentally renames/deletes tmol/kinematics/compiled/compiled_inverse_kin.py into a .cmake query file; commits .claude/settings.local.json, CMake query artifacts, and a generated tmol_primitives.metallib; changes the package repository URL to the contributor fork; enables MPS based on runtime hardware availability rather than an explicit tested build target; and has no current macOS MPS build/runtime lane. It also conflicts with the current CPU/CUDA and PyTorch 2.13/2.14 C++ standard/toolchain work. This needs a clean, narrowly scoped successor with generated artifacts excluded and three independent lanes: Linux CPU regression, Linux CUDA regression, and macOS arm64 MPS compile+forward/backward parity. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This is an MPS fork I needed for a Proteina-Complexa fork running on Apple Silicon.
Florian