2.7.0 Update - CPU only

abs.yaml

@@ -1,7 +1,6 @@
-build_env_vars:
-  ANACONDA_ROCKET_ENABLE_PY313 : yes
-
 # macOS 12.3 or above is required for running the GPU variant (MPS support). No way to specify this for only the GPU
 # variant, so it's specified for both.
 extra_labels_for_os:
   osx-arm64: [ventura]
+
+task_timeout: 72000

recipe/bld.bat

@@ -51,7 +51,7 @@ set DISTUTILS_USE_SDK=1
 set BUILD_TEST=0
 set INSTALL_TEST=0
 :: Don't increase MAX_JOBS to NUMBER_OF_PROCESSORS, as it will run out of heap
-set CPU_COUNT=1
+set CPU_COUNT=4
 set MAX_JOBS=%CPU_COUNT%
 :: Use our Pybind11, Eigen
 set USE_SYSTEM_PYBIND11=1

recipe/conda_build_config.yaml

@@ -1,7 +1,7 @@
 gpu_variant:
   - cpu
   - metal                    # [(osx and arm64)]
-  - cuda-12                  # [(linux and x86_64)]
+#  - cuda-12                  # [(linux and x86_64)]
 c_compiler_version:      # [osx]
   - 17                   # [osx]
 cxx_compiler_version:    # [osx]
@@ -12,10 +12,10 @@ cxx_compiler_version:    # [osx]
 # However there's currently a conda-forge bug that prevents this: https://github.com/conda/conda-build/issues/5048
 MACOSX_SDK_VERSION:          # [(osx and arm64)]
   - 11.1                     # [(osx and arm64)]
-  #- 13.3                     # [(osx and arm64)]
+  - 13.3                     # [(osx and arm64)]
 CONDA_BUILD_SYSROOT:         # [(osx and arm64)]
   - /Library/Developer/CommandLineTools/SDKs/MacOSX11.1.sdk  # [(osx and arm64)]
-  #- /Library/Developer/CommandLineTools/SDKs/MacOSX13.3.sdk  # [(osx and arm64)]
+  - /Library/Developer/CommandLineTools/SDKs/MacOSX13.3.sdk  # [(osx and arm64)]
 zip_keys:                    # [(osx and arm64)]
   - gpu_variant              # [(osx and arm64)]
   - MACOSX_SDK_VERSION       # [(osx and arm64)]
@@ -27,8 +27,8 @@ zip_keys:                    # [(osx and arm64)]
 # Conda-forge didn't do a "megabuild" on osx because it pushed their CI runners over their 6-hour limit. We don't have
 # such a limit.
 megabuild:
-- true
-#- false     # [osx]
+# - true
+- false
 
 # The version of python to use when building libtorch in a "megabuild"
 megabuild_python:

recipe/meta.yaml

@@ -1,8 +1,8 @@
-{% set version = "2.6.0" %}
-{% set sha256 = "3005690eb7b083c443a38c7657938af63902f524ad87a6c83f1aca38c77e3b57" %}
+{% set version = "2.7.0" %}
+{% set sha256 = "ecca266fa2de4235a9fd5a18a33299a9de55ab3babb87f8c297c1c9ab8d436bd" %}
 # Set the RC number to build release candidates. Set to None otherwise
 {% set rc = None %}
-{% set build = 6 %}
+{% set build = 0 %}
 
 # Keep this in sync with the release
 {% set smoke_test_commit = "1eba9b3aa3c43f86f4a2c807ac8e12c4a7767340" %}
@@ -30,8 +30,14 @@ package:
 
 source:
 {% if rc != None %}
-  - git_url: https://github.com/pytorch/pytorch.git
-    git_rev: v{{ version }}-rc{{ rc }}
+  # - git_url: https://github.com/pytorch/pytorch.git
+  #   git_rev: v{{ version.replace(".rc", "-rc") }}
+  # we cannot apply patches to submodules when checking out with git_url, because
+  # then conda switches the patch-application to use git, which cannot construct
+  # a usable ancestor from outside the submodule; the only option then is to
+  # pull in the submodules separately.
+  - url: https://github.com/pytorch/pytorch/archive/refs/tags/v{{ version }}.tar.gz
+    sha256: 04ae0a8babdc9cb9dfc4f8746b2b8aa0f8ed0f9e92835cc4af0bcb01e3969e51
 {% else %}
   # The "pytorch-v" tarballs contain submodules; the "pytorch-" ones don't.
   - url: https://github.com/pytorch/pytorch/releases/download/v{{ version }}/pytorch-v{{ version }}.tar.gz
@@ -42,18 +48,17 @@ source:
       - patches/0003-Force-usage-of-python-3-and-error-without-numpy.patch
       # https://github.com/pytorch/pytorch/pull/137084
       - patches/0004-Help-find-numpy.patch
-      # sympy 1.13.2 was reported to result in test failures on Windows and mac - skipping to 1.13.3 for stability
-      # https://github.com/pytorch/pytorch/pull/133235
-      - patches/0006-Update-sympy-version.patch
+      # https://github.com/pytorch/pytorch/issues/150918 - continue tests on failure due to flaky tests
       - patches/0007-continue-tests-on-failure.patch
       - patches/0008-add-missing-includes.patch
       - patches/0009-use-prefix-include-for-inductor.patch
       - patches/0010-make-ATEN_INCLUDE_DIR-relative-to-TORCH_INSTALL_PREF.patch 
       - patches/0011-remove-DESTINATION-lib-from-CMake-install-TARGETS-di.patch             # [win]
       - patches_submodules/0001-remove-DESTINATION-lib-from-CMake-install-directives.patch  # [win]
-      - patches/0013-simplify-torch.utils.cpp_extension.include_paths-use.patch
       - patches/0014-point-include-paths-to-PREFIX-include.patch
       - patches/0015-point-lib-paths-to-PREFIX-lib.patch
+      - patches/0016-Do-not-check-out-nccl-when-not-building-it.patch
+      - patches/0017-mkl-version-mismatch.patch
 {% endif %}
   - url: https://raw.githubusercontent.com/pytorch/pytorch/{{ smoke_test_commit }}/.ci/pytorch/smoke_test/smoke_test.py
     folder: smoke_test
@@ -64,7 +69,7 @@ build:
   number: {{ build }}
   string: gpu_cuda{{ cuda_compiler_version | replace('.', '') }}_h{{ PKG_HASH }}_{{ PKG_BUILDNUM }}  # [gpu_variant == "cuda-12"]
   string: gpu_mps_h{{PKG_HASH}}_{{ PKG_BUILDNUM }}                                                   # [gpu_variant == "metal"]
-  string: cpu_{{ blas_impl }}_h{{ PKG_HASH }}_{{ PKG_BUILDNUM }}                                 # [gpu_variant == "cpu"]
+  string: cpu_{{ blas_impl }}_h{{ PKG_HASH }}_{{ PKG_BUILDNUM }}                                     # [gpu_variant == "cpu"]
   detect_binary_files_with_prefix: false
   run_exports:
     - {{ pin_subpackage('libtorch', max_pin='x.x') }}
@@ -87,10 +92,10 @@ requirements:
     - cross-python_{{ target_platform }}     # [build_platform != target_platform]
     - numpy  *                               # [megabuild and build_platform != target_platform]
     - numpy                                  # [not megabuild and build_platform != target_platform]
-    #- {{ stdlib('c') }}
+    - {{ stdlib('c') }}
     - {{ compiler('c') }}
     - {{ compiler('cxx') }}
-    - {{ compiler('cuda') }}                 # [(gpu_variant or "").startswith("cuda")]
+    - {{ compiler('cuda') }}                # [(gpu_variant or "").startswith("cuda")]
     - nvtx-c                                 # [cuda_compiler_version != "None" and build_platform != target_platform]
     {% if cuda_major >= 12 %}
     - cuda-driver-dev                        # [build_platform != target_platform]
@@ -114,6 +119,7 @@ requirements:
     # This has a strong run_export so we don't need to put it in `host` or `run`
     # We use llvm-openmp for openblas variants on osx.
     - llvm-openmp 17              # [osx and not (blas_impl == "mkl")]
+    - libcxx 17 # [osx] 
     - libuv     # [win]
     - cmake
     - ninja-base
@@ -158,7 +164,7 @@ requirements:
     - requests
     - future  # [py<313]
     - six
-    - mkl-devel {{ mkl }}           # [blas_impl == "mkl"]
+    - mkl-devel {{ mkl }}          # [blas_impl == "mkl"]
     - openblas-devel {{ openblas }}   # [blas_impl == "openblas"]
     # - libcblas * *_mkl      # [blas_impl == "mkl"]
     # - libcblas              # [blas_impl != "mkl"]
@@ -167,7 +173,7 @@ requirements:
     # We pull in the same versions of mkl and intel-openmp: intel aligns the versions
     # We use intel-openmp for all mkl variants.
     # For openblas on win and linux, we don't specify any openmp implementation; it comes from the compiler.
-    - intel-openmp   {{ mkl }}        # [blas_impl == "mkl"]
+    - intel-openmp {{ mkl }}         # [blas_impl == "mkl"]
     - llvm-openmp 17              # [osx and not (blas_impl == "mkl")]
     - libabseil
     - libprotobuf {{ libprotobuf }}
@@ -183,8 +189,8 @@ requirements:
   # satisfy overlinking checks
   run:
     - {{ pin_compatible('intel-openmp') }}   # [blas_impl == "mkl"]
-    - libuv  # [win]
-    - {{ pin_compatible('magma') }}                       # [(gpu_variant or "").startswith("cuda")]
+    - libuv                                  # [win]
+    - {{ pin_compatible('magma') }}          # [(gpu_variant or "").startswith("cuda")]
 
 # these tests are for the libtorch output below, but due to
 # a particularity of conda-build, that output is defined in
@@ -227,7 +233,6 @@ outputs:
         - "**/shm.dll"                # [win]
         - "**/torch_cpu.dll"          # [win]
         - "**/torch_python.dll"       # [win]
-        - $RPATH/ld64.so.1  # [s390x]
         # libcuda.so is the cuda driver API library and is a system library.
         - "**/libcuda.so*"            # [(gpu_variant or "").startswith("cuda")]
   - name: pytorch
@@ -247,7 +252,6 @@ outputs:
         - "**/shm.dll"                # [win]
         - "**/torch_cpu.dll"          # [win]
         - "**/torch_python.dll"       # [win]
-        - $RPATH/ld64.so.1  # [s390x]
       detect_binary_files_with_prefix: false
       run_exports:
         - {{ pin_subpackage('pytorch', max_pin='x.x') }}
@@ -261,7 +265,7 @@ outputs:
         - python                                 # [build_platform != target_platform]
         - cross-python_{{ target_platform }}     # [build_platform != target_platform]
         - numpy                                  # [build_platform != target_platform]
-        #- {{ stdlib('c') }}
+        - {{ stdlib('c') }}
         - {{ compiler('c') }}
         - {{ compiler('cxx') }}
         - {{ compiler('cuda') }}                 # [(gpu_variant or "").startswith("cuda")]
@@ -288,6 +292,7 @@ outputs:
         # This has a strong run_export so we don't need to put it in `host` or `run`
         # We use llvm-openmp for openblas variants on osx.
         - llvm-openmp 17              # [osx and not (blas_impl == "mkl")]
+        - libcxx 17 # [osx]
         - cmake
         - ninja-base
         # Keep libprotobuf here so that a compatibile version
@@ -327,7 +332,7 @@ outputs:
         - requests
         - future  # [py<313]
         - six
-        - mkl-devel {{ mkl }}           # [blas_impl == "mkl"]
+        - mkl-devel {{ mkl }}             # [blas_impl == "mkl"]
         - openblas-devel {{ openblas }}   # [blas_impl == "openblas"]
         # - libcblas * *_mkl      # [blas_impl == "mkl"]
         # - libcblas              # [blas_impl != "mkl"]
@@ -336,7 +341,7 @@ outputs:
         # We pull in the same versions of mkl and intel-openmp: intel aligns the versions
         # We use intel-openmp for all mkl variants.
         # For openblas on win and linux, we don't specify any openmp implementation; it comes from the compiler.
-        - intel-openmp   {{ mkl }}        # [blas_impl == "mkl"]
+        - intel-openmp {{ mkl }}      # [blas_impl == "mkl"]
         - llvm-openmp 17              # [osx and not (blas_impl == "mkl")]
         - libabseil
         - libprotobuf {{ libprotobuf }}
@@ -374,7 +379,7 @@ outputs:
         - nomkl                 # [blas_impl != "mkl"]
         - fsspec
         # Required to support torch.compile. This is tested in smoke_test.py, which is required to pass
-        - triton 3.1.0       # [(gpu_variant or "").startswith("cuda") and (linux and x86_64)]
+        - triton 3.3.0       # [(gpu_variant or "").startswith("cuda") and (linux and x86_64)]
         # avoid that people without GPUs needlessly download ~0.5-1GB
         # The CUDA version constraint is handled in cuda-version as a run_constrained.
         # However, that doesn't enforce that the package requires a GPU; that needs to be done here.
@@ -410,6 +415,7 @@ outputs:
         - pybind11
         # the inductor "test_aoti_eager..." tests require objcopy
         - binutils  # [linux]
+        - libcxx 17 # [osx]
       imports:
         - torch
       source_files:
@@ -452,7 +458,14 @@ outputs:
         # Note that the `|| true` expression will make the build continue even if the whole script falls over completely
         # (for example, in the case of missing imports). There doesn't seem to be a way of making a script exception return
         # non-zero but failing tests return zero.
-        - python ./test/run_test.py --core --continue-through-error || true
+        # ------------------------------------------------------------------------------------------------
+        # Exclude complex tests that are known to be flaky for -k "not (complex and (linalg_vecdot or dot or vdot))"
+        # https://github.com/pytorch/pytorch/issues/150918
+        - python ./test/run_test.py --core --continue-through-error -k "not (complex and (linalg_vecdot or dot or vdot))" || true # [not win]
+        # lgamma or mvlgamma or multigammaln or gammaln all have these issues on a combination of Intel Xeon processors and Windows Server differences.
+        # enabling these tests on windows will cause numerical differences in the test suite.
+        # This is a non-deterministic issue where between 80-110 tests fail. This has been observed between Pytorch 2.5 and above.
+        - python ./test/run_test.py --core --continue-through-error -k "not ((complex and (linalg_vecdot or dot or vdot)) or lgamma or mvlgamma or multigammaln or gammaln)" || exit 0 # [win]
         # The inductor tests test the torch.compile backend. Using the options below avoids running distributed tests,
         # which would be run if we used the --inductor option. (Distributed tests would only be correctly run on a multi-gpu test platform,
         # which we don't have.)

recipe/patches/0001-windows-FindMKL-add-library-suffix.patch

@@ -8,9 +8,11 @@ This is required because our mdl-devel package contains libraries named like
 
 Index: pytorch/cmake/Modules/FindMKL.cmake
 ===================================================================
---- pytorch.orig/cmake/Modules/FindMKL.cmake	2024-11-19 16:11:16.666292111 -0600
-+++ pytorch/cmake/Modules/FindMKL.cmake	2024-11-19 16:14:15.719045239 -0600
-@@ -119,6 +119,9 @@
+diff --git cmake/Modules/FindMKL.cmake cmake/Modules/FindMKL.cmake
+index a9276de..d340c41 100644
+--- a/cmake/Modules/FindMKL.cmake
++++ b/cmake/Modules/FindMKL.cmake
+@@ -119,6 +119,9 @@ ELSE(WIN32)
      ELSE()
        SET(mklthreads "mkl_intel_thread")
        SET(mklrtls "iomp5" "guide")
@@ -20,7 +22,7 @@ Index: pytorch/cmake/Modules/FindMKL.cmake
      ENDIF()
      SET(mklifaces  "intel")
    ENDIF (CMAKE_COMPILER_IS_GNUCC)
-@@ -268,7 +271,7 @@
+@@ -268,7 +271,7 @@ MACRO(CHECK_ALL_LIBRARIES LIBRARIES OPENMP_TYPE OPENMP_LIBRARY _name _list _flag
            ENDIF(OPENMP_FOUND)
          ELSEIF(${_library} MATCHES "iomp")
            SET(_openmp_type "Intel")
@@ -29,7 +31,7 @@ Index: pytorch/cmake/Modules/FindMKL.cmake
            SET(_openmp_library "${${_prefix}_${_library}_LIBRARY}")
          ELSE()
            MESSAGE(FATAL_ERROR "Unknown OpenMP flavor: ${_library}")
-@@ -421,23 +424,23 @@
+@@ -421,23 +424,23 @@ IF (MKL_LIBRARIES)
    FOREACH(mkl64 ${mkl64s} "_core" "")
      FOREACH(mkls ${mklseq} "")
        IF (NOT MKL_LAPACK_LIBRARIES)

recipe/patches/0002-swap-openmp-search-precedence.patch

@@ -1,8 +1,10 @@
 Index: pytorch/cmake/Modules/FindMKL.cmake
 ===================================================================
---- pytorch.orig/cmake/Modules/FindMKL.cmake	2024-11-19 16:17:57.874552303 -0600
-+++ pytorch/cmake/Modules/FindMKL.cmake	2024-11-19 16:18:29.915971379 -0600
-@@ -108,8 +108,8 @@
+diff --git cmake/Modules/FindMKL.cmake cmake/Modules/FindMKL.cmake
+index d340c41..3eab320 100644
+--- a/cmake/Modules/FindMKL.cmake
++++ b/cmake/Modules/FindMKL.cmake
+@@ -108,8 +108,8 @@ ELSE(WIN32)
        SET(mklthreads "mkl_tbb_thread")
        SET(mklrtls "tbb")
      ELSE()

recipe/patches/0003-Force-usage-of-python-3-and-error-without-numpy.patch

@@ -9,9 +9,11 @@ Subject: [PATCH] Force usage of python 3 and error without numpy
 
 Index: pytorch/cmake/Dependencies.cmake
 ===================================================================
---- pytorch.orig/cmake/Dependencies.cmake	2024-11-12 20:18:55.008695666 -0600
-+++ pytorch/cmake/Dependencies.cmake	2024-11-19 16:18:44.359469010 -0600
-@@ -861,9 +861,9 @@
+diff --git cmake/Dependencies.cmake cmake/Dependencies.cmake
+index 1813f44..36b507f 100644
+--- a/cmake/Dependencies.cmake
++++ b/cmake/Dependencies.cmake
+@@ -848,9 +848,9 @@ if(BUILD_PYTHON)
    if(USE_NUMPY)
      list(APPEND PYTHON_COMPONENTS NumPy)
    endif()
@@ -23,7 +25,7 @@ Index: pytorch/cmake/Dependencies.cmake
  endif()
 
  if(NOT Python_Interpreter_FOUND)
-@@ -880,7 +880,7 @@
+@@ -867,7 +867,7 @@ if(BUILD_PYTHON)
    if(Python_Development.Module_FOUND)
      if(USE_NUMPY)
        if(NOT Python_NumPy_FOUND)

recipe/patches/0004-Help-find-numpy.patch

@@ -9,9 +9,11 @@ Subject: [PATCH] Help find numpy
 
 Index: pytorch/tools/setup_helpers/cmake.py
 ===================================================================
---- pytorch.orig/tools/setup_helpers/cmake.py	2024-11-12 20:18:55.448001614 -0600
-+++ pytorch/tools/setup_helpers/cmake.py	2024-11-19 16:18:54.574972542 -0600
-@@ -305,9 +305,15 @@
+diff --git tools/setup_helpers/cmake.py tools/setup_helpers/cmake.py
+index 84e4dad..8ce7272 100644
+--- a/tools/setup_helpers/cmake.py
++++ b/tools/setup_helpers/cmake.py
+@@ -306,9 +306,15 @@ class CMake:
              sys.exit(1)
          build_options.update(cmake__options)
 

recipe/patches/0007-continue-tests-on-failure.patch

@@ -1,8 +1,10 @@
 Index: pytorch/test/run_test.py
 ===================================================================
---- pytorch.orig/test/run_test.py	2024-11-19 16:09:45.970287774 -0600
-+++ pytorch/test/run_test.py	2024-11-19 16:19:29.415077042 -0600
-@@ -1044,7 +1044,7 @@
+diff --git test/run_test.py test/run_test.py
+index a508d8d..02d9af9 100755
+--- a/test/run_test.py
++++ b/test/run_test.py
+@@ -1384,7 +1384,7 @@ def get_pytest_args(options, is_cpp_test=False, is_distributed_test=False):
      else:
          # When under the normal mode, retry a failed test 2 more times. -x means stop at the first
          # failure

recipe/patches/0008-add-missing-includes.patch

@@ -4,8 +4,10 @@ It won't compile for osx-arm64 metal; unsure why it can on linux-64, but in any
 case, they should be present.
 Index: pytorch/torch/csrc/distributed/c10d/control_plane/Handlers.hpp
 ===================================================================
---- pytorch.orig/torch/csrc/distributed/c10d/control_plane/Handlers.hpp	2024-11-12 20:18:55.762951714 -0600
-+++ pytorch/torch/csrc/distributed/c10d/control_plane/Handlers.hpp	2024-11-19 16:23:52.134859846 -0600
+diff --git torch/csrc/distributed/c10d/control_plane/Handlers.hpp torch/csrc/distributed/c10d/control_plane/Handlers.hpp
+index 70333a3..aa9619e 100644
+--- a/torch/csrc/distributed/c10d/control_plane/Handlers.hpp
++++ b/torch/csrc/distributed/c10d/control_plane/Handlers.hpp
 @@ -4,6 +4,8 @@
  #include <map>
  #include <string>