Remove pinning against torch version

.github/scripts/unittest-linux/run_test.sh

@@ -27,6 +27,7 @@ fi
     export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_unidecode=true
     export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_inflect=true
     export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_pytorch_lightning=true
+    export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MULTIGPU_CUDA=true
     cd test
     pytest torchaudio_unittest -k "not torchscript and not fairseq and not demucs ${PYTEST_K_EXTRA}" 
 )

.github/workflows/build-wheels-m1.yml

@@ -23,12 +23,12 @@ permissions:
 
 jobs:
   generate-matrix:
-    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
+    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@release/2.10
     with:
       package-type: wheel
       os: macos-arm64
       test-infra-repository: pytorch/test-infra
-      test-infra-ref: main
+      test-infra-ref: release/2.10
   build:
     needs: generate-matrix
     strategy:
@@ -39,12 +39,12 @@ jobs:
             smoke-test-script: test/smoke_test/smoke_test_no_ffmpeg.py
             package-name: torchaudio
     name: ${{ matrix.repository }}
-    uses: pytorch/test-infra/.github/workflows/build_wheels_macos.yml@main
+    uses: pytorch/test-infra/.github/workflows/build_wheels_macos.yml@release/2.10
     with:
       repository: ${{ matrix.repository }}
       ref: ""
       test-infra-repository: pytorch/test-infra
-      test-infra-ref: main
+      test-infra-ref: release/2.10
       build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
       smoke-test-script: ${{ matrix.smoke-test-script }}
       runner-type: macos-m1-stable

.github/workflows/build_wheels_aarch64_linux.yml

@@ -23,12 +23,12 @@ permissions:
 
 jobs:
   generate-matrix:
-    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
+    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@release/2.10
     with:
       package-type: wheel
       os: linux-aarch64
       test-infra-repository: pytorch/test-infra
-      test-infra-ref: main
+      test-infra-ref: release/2.10
       with-cuda: enable
   build:
     needs: generate-matrix
@@ -40,12 +40,12 @@ jobs:
             smoke-test-script: test/smoke_test/smoke_test.py
             package-name: torchaudio
     name: ${{ matrix.repository }}
-    uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@main
+    uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@release/2.10
     with:
       repository: ${{ matrix.repository }}
       ref: ""
       test-infra-repository: pytorch/test-infra
-      test-infra-ref: main
+      test-infra-ref: release/2.10
       build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
       pre-script: ${{ matrix.pre-script }}
       post-script: ${{ matrix.post-script }}

.github/workflows/build_wheels_linux.yml

@@ -23,12 +23,12 @@ permissions:
 
 jobs:
   generate-matrix:
-    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
+    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@release/2.10
     with:
       package-type: wheel
       os: linux
       test-infra-repository: pytorch/test-infra
-      test-infra-ref: main
+      test-infra-ref: release/2.10
       with-xpu: enable
   build:
     needs: generate-matrix
@@ -40,12 +40,12 @@ jobs:
             smoke-test-script: test/smoke_test/smoke_test_no_ffmpeg.py
             package-name: torchaudio
     name: ${{ matrix.repository }}
-    uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@main
+    uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@release/2.10
     with:
       repository: ${{ matrix.repository }}
       ref: ""
       test-infra-repository: pytorch/test-infra
-      test-infra-ref: main
+      test-infra-ref: release/2.10
       build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
       smoke-test-script: ${{ matrix.smoke-test-script }}
       package-name: ${{ matrix.package-name }}

.github/workflows/build_wheels_windows.yml

@@ -19,12 +19,12 @@ permissions:
 
 jobs:
   generate-matrix:
-    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
+    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@release/2.10
     with:
       package-type: wheel
       os: windows
       test-infra-repository: pytorch/test-infra
-      test-infra-ref: main
+      test-infra-ref: release/2.10
       with-xpu: enable
   build:
     needs: generate-matrix
@@ -38,12 +38,12 @@ jobs:
             smoke-test-script: test/smoke_test/smoke_test_no_ffmpeg.py
             package-name: torchaudio
     name: ${{ matrix.repository }}
-    uses: pytorch/test-infra/.github/workflows/build_wheels_windows.yml@main
+    uses: pytorch/test-infra/.github/workflows/build_wheels_windows.yml@release/2.10
     with:
       repository: ${{ matrix.repository }}
       ref: ""
       test-infra-repository: pytorch/test-infra
-      test-infra-ref: main
+      test-infra-ref: release/2.10
       build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
       env-script: ${{ matrix.env-script }}
       package-name: ${{ matrix.package-name }}

.github/workflows/integration-test.yml

@@ -13,7 +13,7 @@ jobs:
         python-version: ["3.11"]
         ffmpeg-version: ["7"]
       fail-fast: false
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@release/2.10
     permissions:
       id-token: write
       contents: read

.github/workflows/unittest-linux-cpu.yml

@@ -16,7 +16,7 @@ jobs:
         python-version: ["3.11"]
         ffmpeg-version: ["7"]
       fail-fast: false
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@release/2.10
     permissions:
       id-token: write
       contents: read
@@ -92,5 +92,7 @@ jobs:
         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_FFMPEG=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MULTIGPU_CUDA=true
+
         pytest test/torchaudio_unittest -k "not torchscript and not fairseq and not demucs" -x
         echo "::endgroup::"

.github/workflows/unittest-linux-gpu.yml

@@ -17,7 +17,7 @@ jobs:
         cuda-version: ["12.8"]
         ffmpeg-version: ["7"]
       fail-fast: false
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@release/2.10
     permissions:
       id-token: write
       contents: read
@@ -44,6 +44,8 @@ jobs:
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX_ENCODER=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_FFMPEG=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_demucs=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MULTIGPU_CUDA=true
+
         export CUBLAS_WORKSPACE_CONFIG=:16:8
 
         # Set UPLOAD_CHANNEL

.github/workflows/unittest-macos-cpu.yml

@@ -16,7 +16,7 @@ jobs:
         python-version: ["3.11"]
         ffmpeg-version: ["7"]
       fail-fast: false
-    uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/macos_job.yml@release/2.10
     permissions:
       id-token: write
       contents: read
@@ -89,6 +89,7 @@ jobs:
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_unidecode=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_inflect=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_pytorch_lightning=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MULTIGPU_CUDA=true
 
         pytest test/torchaudio_unittest -k "not torchscript and not fairseq and not demucs and not librosa" -x
         echo "::endgroup::"

.github/workflows/unittest-windows-cpu.yml

@@ -11,7 +11,7 @@ on:
 
 jobs:
   build:
-    uses: pytorch/test-infra/.github/workflows/windows_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/windows_job.yml@release/2.10
     with:
       repository: pytorch/audio
       runner: windows.4xlarge
@@ -54,6 +54,7 @@ jobs:
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_inflect=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_pytorch_lightning=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_sentencepiece=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MULTIGPU_CUDA=true
 
         .github/scripts/unittest-windows/setup_env.sh
         .github/scripts/unittest-windows/install.sh

.github/workflows/unittest-windows-gpu.yml

@@ -11,7 +11,7 @@ on:
 
 jobs:
   build:
-    uses: pytorch/test-infra/.github/workflows/windows_job.yml@main
+    uses: pytorch/test-infra/.github/workflows/windows_job.yml@release/2.10
     with:
       repository: pytorch/audio
       runner: windows.g5.4xlarge.nvidia.gpu
@@ -55,6 +55,7 @@ jobs:
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_inflect=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_pytorch_lightning=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_sentencepiece=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MULTIGPU_CUDA=true
 
         .github/scripts/unittest-windows/setup_env.sh
         .github/scripts/unittest-windows/install.sh

setup.py

@@ -86,8 +86,12 @@ def _main():
     print("-- Git branch:", branch)
     print("-- Git SHA:", sha)
     print("-- Git tag:", tag)
-    pytorch_package_dep = _get_pytorch_version()
-    print("-- PyTorch dependency:", pytorch_package_dep)
+    # This used to be passed to install_requires
+    # which would cause pinning against a specific torch version in releases.
+    # I don't think we want to pin at all?
+    # TODO: revisit if needed. Maybe it's needed for nightlies. Unsure.
+    # pytorch_package_dep = _get_pytorch_version()
+    # print("-- PyTorch dependency:", pytorch_package_dep)
     version = _get_version(sha)
     print("-- Building version", version)
 
@@ -135,7 +139,7 @@ def _main():
             "build_ext": setup_helpers.CMakeBuild,
             "clean": clean,
         },
-        install_requires=[pytorch_package_dep],
+        install_requires=[],
         zip_safe=False,
     )
 

src/libtorchaudio/forced_align/gpu/compute.cu

@@ -122,6 +122,7 @@ void forced_align_impl(
     const int64_t blank,
     Tensor& paths) {
   auto device_index = logProbs.get_device_index();
+  const torch::stable::accelerator::DeviceGuard device_guard(device_index);
   auto defaultStream = libtorchaudio::cuda::getCurrentCUDAStream(device_index);
   auto cpuDataTranferStream = libtorchaudio::cuda::getStreamFromPool(false, device_index);
   const scalar_t kNegInfinity = -std::numeric_limits<scalar_t>::infinity();

src/libtorchaudio/lfilter.cpp

@@ -24,12 +24,10 @@ void host_lfilter_core_loop(
   int64_t n_samples_input = input_signal_windows.size(2);
   int64_t n_samples_output = padded_output_waveform.size(2);
   int64_t n_order = a_coeff_flipped.size(1);
-  scalar_t* output_data =
-      reinterpret_cast<scalar_t*>(padded_output_waveform.data_ptr());
-  const scalar_t* input_data =
-      reinterpret_cast<scalar_t*>(input_signal_windows.data_ptr());
+  scalar_t* output_data = padded_output_waveform.mutable_data_ptr<scalar_t>();
+  const scalar_t* input_data = input_signal_windows.const_data_ptr<scalar_t>();
   const scalar_t* a_coeff_flipped_data =
-      reinterpret_cast<scalar_t*>(a_coeff_flipped.data_ptr());
+      a_coeff_flipped.const_data_ptr<scalar_t>();
 
   torch::stable::parallel_for(
       0, n_channel * n_batch, 1, [&](int64_t begin, int64_t end) {

src/libtorchaudio/rnnt/cpu/compute.cpp

@@ -1,4 +1,6 @@
 #include <libtorchaudio/rnnt/cpu/cpu_transducer.h>
+#include <libtorchaudio/utils.h>
+
 #include <torch/csrc/stable/library.h>
 #include <torch/csrc/stable/ops.h>
 #include <torch/csrc/stable/tensor.h>
@@ -73,15 +75,11 @@ std::tuple<Tensor, Tensor> compute(
   STD_TORCH_CHECK(
       blank >= 0 && blank < logits.size(-1),
       "blank must be within [0, logits.shape[-1])");
-
-  auto max_ivalue = [](const Tensor& t) {
-    return reinterpret_cast<int32_t*>(torch::stable::amax(t, {}).data_ptr())[0];
-  };
-
   STD_TORCH_CHECK(
-      logits.size(1) == max_ivalue(logit_lengths), "input length mismatch");
+      logits.size(1) == torchaudio::util::max<int64_t>(logit_lengths),
+      "input length mismatch");
   STD_TORCH_CHECK(
-      logits.size(2) == max_ivalue(target_lengths) + 1,
+      logits.size(2) == torchaudio::util::max<int64_t>(target_lengths) + 1,
       "output length mismatch");
   STD_TORCH_CHECK(
       targets.size(1) + 1 == logits.size(2), "target length mismatch");
@@ -110,14 +108,12 @@ std::tuple<Tensor, Tensor> compute(
       {DtypeWorkspace<float>::ComputeSizeFromOptions(options)},
       ScalarType::Float);
 
-  // TODO: use t.mutable_data_ptr<..>() instead of reinterpret_cast
-  // when stable ABI Tensor supports mutable_data_ptr templates.
   Workspace<float> workspace(
       /*options=*/options,
       /*dtype_data=*/
-      reinterpret_cast<float*>(float_workspace.mutable_data_ptr()),
+      float_workspace.mutable_data_ptr<float>(),
       /*dtype_size=*/float_workspace.numel(),
-      /*int_data=*/reinterpret_cast<int*>(int_workspace.mutable_data_ptr()),
+      /*int_data=*/int_workspace.mutable_data_ptr<int>(),
       /*int_size=*/int_workspace.numel());
 
   THO_DISPATCH_V2(
@@ -126,12 +122,12 @@ std::tuple<Tensor, Tensor> compute(
       AT_WRAP([&] {
         (Compute</*DTYPE=*/scalar_t, /*CAST_DTYPE=*/float>(
             /*workspace=*/workspace,
-            /*logits=*/reinterpret_cast<scalar_t*>(logits.data_ptr()),
-            /*targets=*/reinterpret_cast<int*>(targets.data_ptr()),
-            /*srcLengths=*/reinterpret_cast<int*>(logit_lengths.data_ptr()),
-            /*tgtLengths=*/reinterpret_cast<int*>(target_lengths.data_ptr()),
-            /*costs=*/reinterpret_cast<scalar_t*>(costs.data_ptr()),
-            /*gradients=*/reinterpret_cast<scalar_t*>(gradients.data_ptr())));
+            /*logits=*/logits.const_data_ptr<scalar_t>(),
+            /*targets=*/targets.const_data_ptr<int>(),
+            /*srcLengths=*/logit_lengths.const_data_ptr<int>(),
+            /*tgtLengths=*/target_lengths.const_data_ptr<int>(),
+            /*costs=*/costs.mutable_data_ptr<scalar_t>(),
+            /*gradients=*/gradients.mutable_data_ptr<scalar_t>()));
       }),
       ScalarType::Float,
       ScalarType::Half);

src/libtorchaudio/rnnt/gpu/compute.cu

@@ -106,9 +106,9 @@ std::tuple<Tensor, Tensor> compute(
 
   Workspace<float> workspace(
       /*options=*/options,
-      /*dtype_data=*/reinterpret_cast<float*>(float_workspace.data_ptr()),
+      /*dtype_data=*/float_workspace.mutable_data_ptr<float>(),
       /*dtype_size=*/float_workspace.numel(),
-      /*int_data=*/reinterpret_cast<int*>(int_workspace.data_ptr()),
+      /*int_data=*/int_workspace.mutable_data_ptr<int>(),
       /*int_size=*/int_workspace.numel());
 
   THO_DISPATCH_V2(
@@ -117,12 +117,12 @@ std::tuple<Tensor, Tensor> compute(
       AT_WRAP([&] {
         (Compute</*DTYPE=*/scalar_t, /*CAST_DTYPE=*/float>(
             /*workspace=*/workspace,
-            /*logits=*/reinterpret_cast<scalar_t*>(logits.data_ptr()),
-            /*targets=*/reinterpret_cast<int*>(targets.data_ptr()),
-            /*srcLengths=*/reinterpret_cast<int*>(logit_lengths.data_ptr()),
-            /*tgtLengths=*/reinterpret_cast<int*>(target_lengths.data_ptr()),
-            /*costs=*/reinterpret_cast<scalar_t*>(costs.data_ptr()),
-            /*gradients=*/reinterpret_cast<scalar_t*>(gradients.data_ptr())));
+            /*logits=*/logits.const_data_ptr<scalar_t>(),
+            /*targets=*/targets.const_data_ptr<int>(),
+            /*srcLengths=*/logit_lengths.const_data_ptr<int>(),
+            /*tgtLengths=*/target_lengths.const_data_ptr<int>(),
+            /*costs=*/costs.mutable_data_ptr<scalar_t>(),
+            /*gradients=*/gradients.mutable_data_ptr<scalar_t>()));
       }),
       ScalarType::Float,
       ScalarType::Half);

src/torchaudio/datasets/cmuarctic.py

@@ -1,4 +1,3 @@
-import csv
 import os
 from pathlib import Path
 from typing import Tuple, Union
@@ -128,9 +127,8 @@ def __init__(
                 )
         self._text = os.path.join(self._path, self._folder_text, self._file_text)
 
-        with open(self._text, "r") as text:
-            walker = csv.reader(text)
-            self._walker = list(walker)
+        with open(self._text, "r", newline=None) as text:
+            self._walker = [[line.rstrip("\n")] for line in text.readlines()]
 
     def __getitem__(self, n: int) -> Tuple[Tensor, int, str, str]:
         """Load the n-th sample from the dataset.

test/torchaudio_unittest/common_utils/__init__.py

@@ -18,6 +18,7 @@
     skipIfNoQengine,
     skipIfPy310,
     skipIfRocm,
+    skipIfSingleCuda,
     TempDirMixin,
     TestBaseMixin,
     TorchaudioTestCase,
@@ -65,6 +66,7 @@ def inject_request(self, request):
     "skipIfNoFFmpeg",
     "skipIfNoHWAccel",
     "skipIfPy310",
+    "skipIfSingleCuda",
     "disabledInCI",
     "get_wav_data",
     "normalize_wav",