v2.11.0 WIP

.github/workflows/conda-build.yml

@@ -70,31 +70,31 @@ jobs:
             DOCKER_IMAGE: quay.io/condaforge/linux-anvil-x86_64:alma9
             tools_install_dir: ~/miniforge3
             build_workspace_dir: build_artifacts
-          - CONFIG: linux_aarch64_c_compiler_version13c_stdlib_version2.17channel_targetsconda-forge_maincuda_compiler_version12.9cxx_compiler_version13is_rcFalse
+          - CONFIG: linux_aarch64_arm_variant_typesbsac_compiler_version13c_stdlib_version2.17channel_targetsconda-forge_maincuda_compiler_version12.9cxx_compiler_version13is_rcFalse
             STORE_BUILD_ARTIFACTS: False
             UPLOAD_PACKAGES: True
             os: ubuntu
             runs_on: ['namespace-profile-16cpu-on-linux-64']
             DOCKER_IMAGE: quay.io/condaforge/linux-anvil-x86_64:alma9
             tools_install_dir: ~/miniforge3
             build_workspace_dir: build_artifacts
-          - CONFIG: linux_aarch64_c_compiler_version13c_stdlib_version2.17channel_targetsconda-forge_maincuda_compiler_versionNonecxx_compiler_version13is_rcFalse
+          - CONFIG: linux_aarch64_arm_variant_typesbsac_compiler_version13c_stdlib_version2.17channel_targetsconda-forge_maincuda_compiler_versionNonecxx_compiler_version13is_rcFalse
             STORE_BUILD_ARTIFACTS: False
             UPLOAD_PACKAGES: True
             os: ubuntu
             runs_on: ['namespace-profile-16cpu-on-linux-64']
             DOCKER_IMAGE: quay.io/condaforge/linux-anvil-x86_64:alma9
             tools_install_dir: ~/miniforge3
             build_workspace_dir: build_artifacts
-          - CONFIG: linux_aarch64_c_compiler_version13c_stdlib_version2.28channel_targetsconda-forge_maincuda_compiler_version13.0cxx_compiler_version13is_rcFalse
+          - CONFIG: linux_aarch64_arm_variant_typesbsac_compiler_version13c_stdlib_version2.28channel_targetsconda-forge_maincuda_compiler_version13.0cxx_compiler_version13is_rcFalse
             STORE_BUILD_ARTIFACTS: False
             UPLOAD_PACKAGES: True
             os: ubuntu
             runs_on: ['namespace-profile-16cpu-on-linux-64']
             DOCKER_IMAGE: quay.io/condaforge/linux-anvil-x86_64:alma9
             tools_install_dir: ~/miniforge3
             build_workspace_dir: build_artifacts
-          - CONFIG: linux_aarch64_c_compiler_version14c_stdlib_version2.34channel_targetsconda-forge_maincuda_compiler_version12.9cxx_compiler_version14is_rcFalse
+          - CONFIG: linux_aarch64_arm_variant_typetegrac_compiler_version14c_stdlib_version2.34channel_targetsconda-forge_maincuda_compiler_version12.9cxx_compiler_version14is_rcFalse
             STORE_BUILD_ARTIFACTS: False
             UPLOAD_PACKAGES: True
             os: ubuntu

recipe/build.sh

@@ -16,6 +16,11 @@ export PACKAGE_TYPE=conda
 # remove pyproject.toml to avoid installing deps from pip
 rm -rf pyproject.toml
 
+# remove runtime pin for setuptools, upstream added it to workaround
+# breakage from transitive dependencies using pkg_resources. we can handle
+# these dependencies directly in conda-forge.
+sed -i -e '/setuptools<82/d' setup.py
+
 # uncomment to debug cmake build
 # export CMAKE_VERBOSE_MAKEFILE=1
 

recipe/meta.yaml

@@ -1,6 +1,6 @@
-# if you wish to build release candidate number X, append the version string with ".rcX"
-{% set version = "2.10.0" %}
-{% set build = 4 %}
+# if you wish to build release candidate number X, append the version string with "-rcX"
+{% set version = "2.11.0" %}
+{% set build = 0 %}
 
 # Use a higher build number for the CUDA variant, to ensure that it's
 # preferred by conda's solver, and it's preferentially
@@ -25,11 +25,11 @@ package:
 source:
 {% if "rc" in version %}
   - url: https://download.pytorch.org/source_code/test/pytorch-v{{ version }}.tar.gz
-    sha256: f35b2d7839b284410e5be9ec2eeb7a3049e09c1b8f6a871d3f2cad495d93dcd6
+    sha256: 3035931fff5b79e0300db69b0249e0c9c7ea5b394f451cb80e0280cbc2affcc2
 {% 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
-    sha256: fa8ccbe87f83f48735505371c1c313b4aa6db400b0ae4f8a02844d1e150c695f
+    sha256: ab3fde9e7e382f45ac942be6ea2c2ef362c5ccd6f55ed6d5f35e6ea81d3ab88e
 {% endif %}
     patches:
       - patches/0001-Force-usage-of-python-3-and-error-without-numpy.patch
@@ -50,12 +50,11 @@ source:
       # backport https://github.com/pytorch/pytorch/pull/166824
       - patches/0011-Add-USE_SYSTEM-options-for-KLEIDI-CUDNN_FRONTEND-CUT.patch
       - patches/0012-Fix-building-kineto-against-system-fmt.patch
-      # backport https://github.com/pytorch/pytorch/pull/159828
-      - patches/0013-Attempt-to-fix-torch.backends.cudnn.rnn-import.patch
       - patches/0014-Use-Intel-LLVM-openmp.patch
-      # backport https://github.com/pytorch/pytorch/pull/174647
-      - patches/0015-Fix-ICE-in-GCC-14-with-arm.patch
       - patches/0016-for-win-CUDA-remove-USE_CUDA-guard-for-skip-in-compi.patch         # [cuda_compiler_version != "None"]
+      # backport https://github.com/pytorch/pytorch/pull/175283
+      - patches/0017-Ensure-test_tensorinv-uses-well-conditioned-inputs-1.patch
+      - patches/0018-Declare-_tryToInferTypeImpl-with-TORCH_PYTHON_API.patch
       - patches_submodules/tensorpipe/0001-switch-away-from-find_package-CUDA.patch
 
 build:
@@ -132,13 +131,14 @@ requirements:
     - libcusparse-dev
     - libmagma-devel
     - nccl                  # [linux]
+    - nccl <2.29            # [linux and cuda_compiler_version=="12.9" and arm_variant_type!="tegra"]
     - nvtx-c
     {% endif %}
     # other requirements
     - python 3.12
     - numpy *
     - pip
-    - setuptools
+    - setuptools <82
     - pyyaml
     - requests
     - six
@@ -278,13 +278,14 @@ outputs:
         - libcusparse-dev
         - libmagma-devel
         - nccl                  # [linux]
+        - nccl <2.29            # [linux and cuda_compiler_version=="12.9" and arm_variant_type!="tegra"]
         - nvtx-c
         {% endif %}
         # other requirements
         - python
         - numpy
         - pip
-        - setuptools
+        - setuptools <82
         - pyyaml
         - requests
         - six
@@ -325,11 +326,9 @@ outputs:
         - pybind11
         # https://github.com/pytorch/pytorch/pull/175115
         - pybind11 <3.0.2
-        - setuptools
+        - setuptools <82
         - sympy >=1.13.3
         - typing_extensions >=4.10.0
-        # see https://github.com/conda-forge/sympy-feedstock/issues/67
-        - mpmath <1.4
       run_constrained:
         # https://github.com/conda-forge/pytorch-cpu-feedstock/issues/65
         - pytorch-cpu {{ version }}    # [cuda_compiler_version == "None"]
@@ -517,6 +516,23 @@ outputs:
         {% set skips = skips ~ " or test_qengine" %}                                    # [osx and arm64]
         # flaky failure on osx
         {% set skips = skips ~ " or test_LayerNorm_numeric_mps" %}                      # [osx and arm64]
+        # precision errors
+        {% set skips = skips ~ " or test_forward_nn_Linear" %}                          # [osx and arm64]
+        {% set skips = skips ~ " or test_forward_nn_TransformerEncoderLayer_train_mode_mps_float16" %}  # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_GRUCell_mps" %}      # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_GRU_eval_mode_mps" %}  # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_GRU_train_mode_mps" %}  # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_LSTMCell_mps" %}     # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_Linear_mps" %}       # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_MultiheadAttention_eval_mode_mps_float16" %}  # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_MultiheadAttention_train_mode_mps_float16" %}  # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_RNNCell_mps" %}      # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_RNN_eval_mode_mps" %}  # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_RNN_train_mode_mps" %}  # [osx and arm64]
+        {% set skips = skips ~ " or test_transformerencoderlayer_mps_float32" %}        # [osx and arm64]
+        {% set skips = skips ~ " or test_transformerencoderlayer_gelu_mps_float32" %}   # [osx and arm64]
+        {% set skips = skips ~ " or test_grad_nn_MultiheadAttention_eval_mode_cpu_float64" %}   # [osx and arm64]
+        {% set skips = skips ~ " or test_non_contiguous_tensors_nn_CrossEntropyLoss_mps_float32" %}   # [osx and arm64]
         # some warning-related failure, maybe it's broken by --disable-warnings?
         {% set skips = skips ~ " or test_cpp_warnings_have_python_context_cpu" %}
         {% set skips = skips ~ " or test_cpp_warnings_have_python_context_cuda" %}
@@ -533,15 +549,15 @@ outputs:
         # disable hypothesis because it randomly yields health check errors
 
         # the opengpu server has a card with sm_70, an architecture dropped by CUDA 13.0
-        {% if (cuda_compiler_version or "0").split(".")[0]|int < 13 %}
-        - pytest {{ jobs }} {{ tests }} -k "not ({{ skips }})" -m "not hypothesis" --durations=50 --timeout=1200 --disable-warnings    # [not aarch64 or py==312]
-        {% endif %}
+        - set ONEDNN_VERBOSE=all       # [win]
+        - export ONEDNN_VERBOSE=all    # [unix]
+        - pytest -v {{ jobs }} {{ tests }} -k "not ({{ skips }})" -m "not hypothesis" --durations=50 --timeout=1200 --disable-warnings --force-short-summary    # [not aarch64 or py==312]
 
         # regression test for https://github.com/conda-forge/pytorch-cpu-feedstock/issues/329, where we picked up
         # duplicate `.pyc` files due to newest py-ver (3.13) in the build environment not matching the one in host;
         # obviously this test can only be done for other python versions.
-        - test ! -f $SP_DIR/functorch/__pycache__/__init__.cpython-313.pyc          # [py!=313 and unix]
-        - if exist %SP_DIR%\functorch\__pycache__\__init__.cpython-313.pyc exit 1   # [py!=313 and win]
+        - test ! -f $SP_DIR/functorch/__pycache__/__init__.cpython-314.pyc          # [py!=314 and unix]
+        - if exist %SP_DIR%\functorch\__pycache__\__init__.cpython-314.pyc exit 1   # [py!=314 and win]
 
         # test integrity of CMake metadata and ensure that THPLayoutType is visible as a symbol from libtorch_python
         - cd cmake_test

recipe/patches/0005-use-our-own-PREFIX-for-include-paths-etc.patch

@@ -1,7 +1,7 @@
-From 44659d96e67c15c4b626ae98046898f0de47d93b Mon Sep 17 00:00:00 2001
+From 97ef248e4150e8cace5f21c8f8fa9eb87f768ff0 Mon Sep 17 00:00:00 2001
 From: "H. Vetinari" <h.vetinari@gmx.com>
 Date: Thu, 23 Jan 2025 22:58:14 +1100
-Subject: [PATCH 05/16] use our own PREFIX for include paths etc.
+Subject: [PATCH 05/13] use our own PREFIX for include paths etc.
 MIME-Version: 1.0
 Content-Type: text/plain; charset=UTF-8
 Content-Transfer-Encoding: 8bit
@@ -33,39 +33,40 @@ Co-Authored-By: Daniel Petry <dpetry@anaconda.com>
 Co-Authored-By: Michał Górny <mgorny@quansight.com>
 Co-Authored-By: Tobias Fischer <info@tobiasfischer.info>
 ---
- cmake/TorchConfig.cmake.in     |  7 ++++---
+ cmake/TorchConfig.cmake.in     | 13 ++++++++++---
  torch/_inductor/cpp_builder.py |  4 +++-
  torch/utils/cpp_extension.py   | 34 ++++++++++++++++++++--------------
- 3 files changed, 27 insertions(+), 18 deletions(-)
+ 3 files changed, 33 insertions(+), 18 deletions(-)
 
 diff --git a/cmake/TorchConfig.cmake.in b/cmake/TorchConfig.cmake.in
-index 0b32ffa99ce..dcce7b38015 100644
+index abf5c814911..475dc3b88ce 100644
 --- a/cmake/TorchConfig.cmake.in
 +++ b/cmake/TorchConfig.cmake.in
-@@ -53,14 +53,15 @@ else()
+@@ -53,9 +53,16 @@ else()
  endif()
 
  # Include directories.
--if(EXISTS "${TORCH_INSTALL_PREFIX}/include")
+-set(TORCH_INCLUDE_DIRS
+-  ${TORCH_INSTALL_PREFIX}/include
+-  ${TORCH_INSTALL_PREFIX}/include/torch/csrc/api/include)
 +if(EXISTS "${TORCH_INSTALL_PREFIX}/include/torch/csrc/api/include")
 +  # top-level include directory
-   set(TORCH_INCLUDE_DIRS
--    ${TORCH_INSTALL_PREFIX}/include
-     ${TORCH_INSTALL_PREFIX}/include/torch/csrc/api/include)
- else()
++  set(TORCH_INCLUDE_DIRS
++    ${TORCH_INSTALL_PREFIX}/include/torch/csrc/api/include)
++else()
 +  # site-packages include directory
-   set(TORCH_INCLUDE_DIRS
-     ${TORCH_INSTALL_PREFIX}/include
--    ${TORCH_INSTALL_PREFIX}/include/torch/csrc/api/include)
++  set(TORCH_INCLUDE_DIRS
++    ${TORCH_INSTALL_PREFIX}/include
 +    ${TORCH_INSTALL_PREFIX}/../../../../include/torch/csrc/api/include)
- endif()
++endif()
 
  # Library dependencies.
+ if(@BUILD_SHARED_LIBS@)
 diff --git a/torch/_inductor/cpp_builder.py b/torch/_inductor/cpp_builder.py
-index 6a6b7d15ae3..0a4724e5c17 100644
+index 6dd6e0d2b5c..fbfa3175836 100644
 --- a/torch/_inductor/cpp_builder.py
 +++ b/torch/_inductor/cpp_builder.py
-@@ -1520,10 +1520,12 @@ def get_cpp_torch_options(
+@@ -1522,10 +1522,12 @@ def get_cpp_torch_options(
          + python_include_dirs
          + torch_include_dirs
          + omp_include_dir_paths
@@ -80,10 +81,10 @@ index 6a6b7d15ae3..0a4724e5c17 100644
      passthrough_args = (
          sys_libs_passthrough_args + isa_ps_args_build_flags + omp_passthrough_args
 diff --git a/torch/utils/cpp_extension.py b/torch/utils/cpp_extension.py
-index f29c382f0e3..d865df1684b 100644
+index a63bff50d5e..7da14c2429c 100644
 --- a/torch/utils/cpp_extension.py
 +++ b/torch/utils/cpp_extension.py
-@@ -1567,31 +1567,37 @@ def include_paths(device_type: str = "cpu", torch_include_dirs=True) -> list[str
+@@ -1604,31 +1604,37 @@ def include_paths(device_type: str = "cpu", torch_include_dirs=True) -> list[str
      Returns:
          A list of include path strings.
      """