Merge branch 'fix_v2.4.1' into 'v2.4.1-fastpt'

fix 104994 and add jenkins.sh

See merge request !1

jenkins.sh

+#!/bin/bash
+
+pip install kaldi_io -i https://mirrors.huaweicloud.com/artifactory/pypi-public/simple
+pip install SoundFile -i https://mirrors.huaweicloud.com/artifactory/pypi-public/simple
+pip install numpy==1.24.3 -i https://mirrors.huaweicloud.com/artifactory/pypi-public/simple
+
+set -e
+
+source /opt/dtk/env.sh
+source /usr/local/bin/fastpt -C
+
+echo "开始编译组件torchaudio"
+python3 setup.py -v bdist_wheel
+echo "组件torchaudio 编译完成"
+
+ls dist/
+pip install dist/torchaudio*.whl --force-reinstall --no-deps
+echo "组件torchaudio安装完成"
+
+echo "开始torchaudio单元测试"
+if command -v pytest &> /dev/null; then
+    echo "pytest 已安装，版本: $(pytest --version)"
+    # 跳过安装
+else
+    echo "pytest 未安装，正在安装..."
+    # 使用 pip 安装 (最常见)
+    pip install pytest
+fi
+
+pytest -vs ./test

src/libtorchaudio/forced_align/gpu/compute.cu

@@ -17,7 +17,7 @@ namespace torchaudio {
 namespace alignment {
 namespace gpu {
 template <typename scalar_t, typename target_t>
-__global__ void falign_cuda_step_kernel(
+__global__ __launch_bounds__(1024) void falign_cuda_step_kernel(
     const torch::PackedTensorAccessor32<scalar_t, 3, torch::RestrictPtrTraits>
         logProbs_a,
     const torch::PackedTensorAccessor32<target_t, 2, torch::RestrictPtrTraits>