Merge pull request #27715 from hadim/torchaudio

add torchaudio

recipes/torchaudio/build.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+set -ex
+
+if [[ ${cuda_compiler_version} != "None" ]]; then
+  # Set the CUDA arch list from
+  # https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/build_pytorch.sh
+  if [[ ${cuda_compiler_version} == 11.8 ]]; then
+    export TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6;8.9+PTX"
+    export CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME
+  elif [[ ${cuda_compiler_version} == 12.0 ]]; then
+    export TORCH_CUDA_ARCH_LIST="5.0;6.0;6.1;7.0;7.5;8.0;8.6;8.9;9.0+PTX"
+    # $CUDA_HOME not set in CUDA 12.0. Using $PREFIX
+    export CUDA_TOOLKIT_ROOT_DIR="${PREFIX}"
+    # CUDA_HOME must be set for the build to work in torchaudio
+    export CUDA_HOME="${PREFIX}"
+  else
+    echo "unsupported cuda version. edit build.sh"
+    exit 1
+  fi
+
+  if [[ "${target_platform}" != "${build_platform}" ]]; then
+    export CUDA_TOOLKIT_ROOT=${PREFIX}
+  fi
+
+  export USE_CUDA=1
+  export BUILD_CUDA_CTC_DECODER=1
+else
+  export USE_CUDA=0
+  export BUILD_CUDA_CTC_DECODER=0
+fi
+
+export USE_ROCM=0
+export USE_OPENMP=1
+export BUILD_CPP_TEST=0
+
+# sox is buggy
+export BUILD_SOX=0
+
+# FFMPEG is buggy
+export USE_FFMPEG=0
+# export FFMPEG_ROOT="${PREFIX}"
+
+# RNNT loss is buggy
+export BUILD_RNNT=0
+
+export CMAKE_C_COMPILER="$CC"
+export CMAKE_CXX_COMPILER="$CXX"
+export CMAKE_GENERATOR="Ninja"
+
+python -m pip install . -vv

recipes/torchaudio/patches/0001-point-to-correct-prefix.patch

@@ -0,0 +1,16 @@
+diff --git a/tools/setup_helpers/extension.py b/tools/setup_helpers/extension.py
+index 2415bbae..f724c4b8 100644
+--- a/tools/setup_helpers/extension.py
++++ b/tools/setup_helpers/extension.py
+@@ -124,8 +124,9 @@ class CMakeBuild(build_ext):
+
+         cmake_args = [
+             f"-DCMAKE_BUILD_TYPE={cfg}",
+-            f"-DCMAKE_PREFIX_PATH={torch.utils.cmake_prefix_path}",
+-            f"-DCMAKE_INSTALL_PREFIX={extdir}",
++            f"-DCMAKE_PREFIX_PATH={os.environ['PREFIX']}",
++            f"-DCMAKE_LIBRARY_PREFIX={os.environ['PREFIX']}",
++            f"-DCMAKE_INSTALL_PREFIX={os.environ['SP_DIR']}/torch/lib",
+             "-DCMAKE_VERBOSE_MAKEFILE=ON",
+             f"-DPython_INCLUDE_DIR={distutils.sysconfig.get_python_inc()}",
+             f"-DBUILD_CPP_TEST={'ON' if _BUILD_CPP_TEST else 'OFF'}",

recipes/torchaudio/patches/0002-use-conda-cuda.patch

@@ -0,0 +1,15 @@
+diff --git a/tools/setup_helpers/extension.py b/tools/setup_helpers/extension.py
+index 2415bbaedb..fcf6a69d93 100644
+--- a/tools/setup_helpers/extension.py
++++ b/tools/setup_helpers/extension.py
+@@ -150,10 +150,6 @@ def build_extension(self, ext):
+             _arches = [arch[:-4] if arch.endswith("+PTX") else f"{arch}-real" for arch in _arches]
+             cmake_args += [f"-DCMAKE_CUDA_ARCHITECTURES={';'.join(_arches)}"]
+
+-        if platform.system() != "Windows" and CUDA_HOME is not None:
+-            cmake_args += [f"-DCMAKE_CUDA_COMPILER='{CUDA_HOME}/bin/nvcc'"]
+-            cmake_args += [f"-DCUDA_TOOLKIT_ROOT_DIR='{CUDA_HOME}'"]
+-
+         # Default to Ninja
+         if "CMAKE_GENERATOR" not in os.environ or platform.system() == "Windows":
+             cmake_args += ["-GNinja"]

recipes/torchaudio/patches/0003-Apply-CMAKE_ARGS-if-set.patch

@@ -0,0 +1,22 @@
+diff --git a/tools/setup_helpers/extension.py b/tools/setup_helpers/extension.py
+index 2415bbaedb..2b6243a790 100644
+--- a/tools/setup_helpers/extension.py
++++ b/tools/setup_helpers/extension.py
+@@ -3,6 +3,7 @@
+ import platform
+ import subprocess
+ from pathlib import Path
++import shlex
+
+ import torch
+ from setuptools import Extension
+@@ -141,6 +142,9 @@ def build_extension(self, ext):
+             f"-DUSE_OPENMP:BOOL={'ON' if _USE_OPENMP else 'OFF'}",
+             f"-DUSE_FFMPEG:BOOL={'ON' if _USE_FFMPEG else 'OFF'}",
+         ]
++        if "CMAKE_ARGS" in os.environ:
++            cmake_args += shlex.split(os.environ["CMAKE_ARGS"])
++
+         build_args = ["--target", "install"]
+         # Pass CUDA architecture to cmake
+         if _TORCH_CUDA_ARCH_LIST is not None:

recipes/torchaudio/patches/0004-add-missing-ffmpeg-skip.patch

@@ -0,0 +1,21 @@
+From d7e13dfe2ff32c9a6020bd56e97b1731e508243e Mon Sep 17 00:00:00 2001
+From: Tobias Fischer <[email protected]>
+Date: Sat, 5 Oct 2024 20:52:42 +1000
+Subject: [PATCH] Add missing @skipIfNoFFmpeg for TestFileObject
+
+---
+ test/torchaudio_unittest/backend/dispatcher/ffmpeg/info_test.py | 1 +
+ 1 file changed, 1 insertion(+)
+
+diff --git a/test/torchaudio_unittest/backend/dispatcher/ffmpeg/info_test.py b/test/torchaudio_unittest/backend/dispatcher/ffmpeg/info_test.py
+index 58a085636b..bf93a75289 100644
+--- a/test/torchaudio_unittest/backend/dispatcher/ffmpeg/info_test.py
++++ b/test/torchaudio_unittest/backend/dispatcher/ffmpeg/info_test.py
+@@ -404,6 +404,7 @@ def read(self, n):
+
+
+ @skipIfNoExec("sox")
++@skipIfNoFFmpeg
+ class TestFileObject(FileObjTestBase, PytorchTestCase):
+     _info = partial(get_info_func(), backend="ffmpeg")
+

recipes/torchaudio/recipe.yaml

@@ -0,0 +1,183 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/prefix-dev/recipe-format/main/schema.json
+
+context:
+  name: torchaudio
+  version: "2.4.1"
+  build_number: 0
+  cuda_version: ${{ env.get("CONDA_OVERRIDE_CUDA", default="None") }}
+  cuda: ${{ "true" if cuda_version != "None" else "false" }}
+  cuda_build_string: cuda_${{ cuda_version | version_to_buildstring }}
+  string_prefix: ${{ cuda_build_string if cuda ==  "true" else "cpu_" }}
+
+package:
+  name: ${{ name|lower }}
+  version: ${{ version }}
+
+source:
+  - url: https://github.com/pytorch/audio/archive/refs/tags/v${{ version }}.tar.gz
+    sha256: a6b7e1b606ae353b9c4db1c8e754263f53f0457e9ab538e9669c0af0c46a0d5c
+    patches:
+      - patches/0001-point-to-correct-prefix.patch
+      - patches/0002-use-conda-cuda.patch
+      - patches/0003-Apply-CMAKE_ARGS-if-set.patch
+      - patches/0004-add-missing-ffmpeg-skip.patch
+
+build:
+  number: ${{ build_number }}
+  skip:
+    - win
+  string: ${{ string_prefix }}py${{ python | version_to_buildstring }}h${{ hash }}_${{ build_number }}
+  variant:
+    use_keys:
+      # use cuda from the variant config, e.g. to build multiple CUDA variants
+      - ${{ "cuda" if cuda == "true" }}
+    # this will down-prioritize the cuda variant versus other variants of the package
+    down_prioritize_variant: ${{ 1 if cuda == "true" else 0 }}
+  script:
+    file: build
+    env:
+      cuda_compiler_version: ${{ cuda_version | default('None') }}
+
+requirements:
+  build:
+    - if: build_platform != target_platform
+      then:
+        - python
+        - cross-python_${{ target_platform }}
+        - pytorch
+
+        - if: match(cuda_version, ">=12")
+          then:
+            - cuda-driver-dev
+            - cuda-cudart-dev
+            - cuda-nvrtc-dev
+            - cuda-nvtx-dev
+            - cuda-nvml-dev
+            - cuda-profiler-api
+            - libcublas-dev
+            - libcufft-dev
+            - libcurand-dev
+            - libcusolver-dev
+            - libcusparse-dev
+
+    - ${{ compiler('cxx') }}
+    - ${{ compiler('c') }}
+    - ${{ stdlib("c") }}
+    - cmake
+    - ninja
+    - ccache
+    - git
+
+    - if: cuda == "true"
+      then:
+        - ${{ compiler('cuda') }}
+        - cuda-version ==${{ cuda_version }}
+
+  host:
+    - python
+    - pip
+    - setuptools
+    - pytorch
+    - ${{ "pytorch * cuda*" if cuda == "true" }}
+    - ${{ "pytorch * cpu*" if cuda == "false" }}
+
+    - bzip2
+    - kaldi
+    - pybind11
+    # - sox
+    # - ffmpeg
+    - xz
+    - zlib
+
+    - if: cuda == "true"
+      then:
+        - cuda-version ==${{ cuda_version }}
+
+    - if: match(cuda_version, ">=12")
+      then:
+        - cuda-driver-dev
+        - cuda-cudart-dev
+        - cuda-nvrtc-dev
+        - cuda-nvtx-dev
+        - cuda-nvml-dev
+        - cuda-profiler-api
+        - libcublas-dev
+        - libcufft-dev
+        - libcurand-dev
+        - libcusolver-dev
+        - libcusparse-dev
+  run:
+    - python
+    - numpy
+    - kaldi
+    - ${{ "pytorch * cuda*" if cuda == "true" }}
+    - ${{ "pytorch * cpu*" if cuda == "false" }}
+
+  ignore_run_exports:
+    from_package:
+      - if: match(cuda_version, ">=12")
+        then:
+          - cuda-nvrtc-dev
+          - cuda-nvtx-dev
+          - libcublas-dev
+          - libcufft-dev
+          - libcurand-dev
+          - libcusolver-dev
+          - libcusparse-dev
+
+tests:
+  - python:
+      imports:
+        - torchaudio
+        - torchaudio.backend
+        - torchaudio.compliance
+        - torchaudio.datasets
+        - torchaudio.functional
+        - torchaudio.models
+        - torchaudio.pipelines
+        - torchaudio.kaldi_io
+        - torchaudio.utils
+        - torchaudio.sox_effects
+        - torchaudio.transforms
+      pip_check: true
+
+  - requirements:
+      run:
+        - pytest
+        - scipy
+        - numpy
+        - librosa
+        - expecttest
+        - requests
+        - hypothesis
+        - inflect
+        - kaldi_io
+        - parameterized
+        - pysoundfile
+        - transformers
+        - unidecode
+        - inflect
+        # - sox
+        - pytorch-lightning
+        - sentencepiece
+    files:
+      source:
+        - test/
+        - examples/
+      recipe:
+        - run_tests.sh
+    script:
+      - bash run_tests.sh
+
+about:
+  homepage: https://github.com/pytorch/audio
+  license: BSD-2-Clause
+  license_file:
+    - LICENSE
+    - third_party/LICENSES_BUNDLED.txt
+  summary: Data manipulation and transformation for audio signal processing, powered by PyTorch
+
+extra:
+  recipe-maintainers:
+    - Tobias-Fischer
+    - h-vetinari

recipes/torchaudio/run_tests.sh

@@ -0,0 +1,75 @@
+#!/bin/bash
+set -ex
+
+export CI=true
+
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_APPLY_CMVN_SLIDING="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_FBANK_FEATS="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_KALDI_PITCH_FEATS="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_MFCC_FEATS="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CMD_COMPUTE_SPECTROGRAM_FEATS="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_KALDI="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CUDA="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_HW_ACCEL="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_AUDIO_OUT_DEVICE="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MACOS="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_TEMPORARY_DISABLED="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX_DECODER="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX_ENCODER="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CTC_DECODER="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_demucs="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_fairseq="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_QUANTIZATION="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_RIR="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_FFMPEG="true"
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX="true"
+
+
+## OVERVIEW OF SKIPPED TESTS
+
+# Output 0 of UnbindBackward0 is a view and is being modified inplace. 
+# This view is the output of a function that returns multiple views. 
+# Such functions do not allow the output views to be modified inplace. 
+# You should replace the inplace operation by an out-of-place one.
+tests_to_skip="TestAutogradLfilterCPU"
+tests_to_skip="test_deemphasis or ${tests_to_skip}"
+
+# 'torchaudio' object has no attribute 'rnnt_loss'
+tests_to_skip="rnnt or ${tests_to_skip}"
+
+# 'torchaudio' object has no attribute 'ray_tracing'
+tests_to_skip="ray_tracing or ${tests_to_skip}"
+
+# object has no attribute _simulate_rir:
+tests_to_skip="test_simulate_rir or ${tests_to_skip}"
+
+# ValueError: invalid version number '0.10.2.post1'
+tests_to_skip="test_create_mel or ${tests_to_skip}"
+
+# RuntimeError: torchaudio.functional._alignment.forced_align Requires alignment extension, but TorchAudio is not compiled with it.
+# Please build TorchAudio with alignment support.
+tests_to_skip="test_forced_align or ${tests_to_skip}"
+
+# Very slow on CI:
+tests_to_skip="hubert_large or ${tests_to_skip}"
+tests_to_skip="hubert_xlarge or ${tests_to_skip}"
+tests_to_skip="hubert_pretrain_large or ${tests_to_skip}"
+tests_to_skip="hubert_pretrain_xlarge or ${tests_to_skip}"
+tests_to_skip="wavlm_large or ${tests_to_skip}"
+tests_to_skip="test_masking_iid or ${tests_to_skip}"
+tests_to_skip="test_mvdr_0_ref_channel or ${tests_to_skip}"
+tests_to_skip="test_rtf_mvdr or ${tests_to_skip}"
+tests_to_skip="test_souden_mvdr or ${tests_to_skip}"
+
+# Segfault on CI (probably due to limited memory):
+tests_to_skip="test_pitch_shift_shape_2 or ${tests_to_skip}"
+tests_to_skip="test_paper_configuration or ${tests_to_skip}"
+tests_to_skip="test_oscillator_bank or ${tests_to_skip}"
+tests_to_skip="test_PitchShift or ${tests_to_skip}"
+tests_to_skip="test_pitch_shift_resample_kernel or ${tests_to_skip}"
+tests_to_skip="test_quantize_torchscript_1_wav2vec2_large or ${tests_to_skip}"
+tests_to_skip="test_quantize_torchscript_2_wav2vec2_large_lv60k or ${tests_to_skip}"
+
+
+pytest -v test/torchaudio_unittest/ -k "not (${tests_to_skip})"