2025-10-09 nightly release (986f10c)

Author: pytorchbot

.github/workflows/build_ffmpeg.yaml

@@ -27,7 +27,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        ffmpeg-version: ["4.4.4", "5.1.4", "6.1.1", "7.0.1"]
+        ffmpeg-version: ["4.4.4", "5.1.4", "6.1.1", "7.0.1", "8.0"]
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write
@@ -52,7 +52,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        ffmpeg-version: ["4.4.4", "5.1.4", "6.1.1", "7.0.1"]
+        ffmpeg-version: ["4.4.4", "5.1.4", "6.1.1", "7.0.1", "8.0"]
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     with:
       job-name: Build
@@ -75,7 +75,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        ffmpeg-version: ["4.4.4", "5.1.4", "6.1.1", "7.0.1"]
+        ffmpeg-version: ["4.4.4", "5.1.4", "6.1.1", "7.0.1", "8.0"]
     uses: pytorch/test-infra/.github/workflows/windows_job.yml@main
     with:
       job-name: Build

packaging/build_ffmpeg.sh

@@ -104,6 +104,14 @@ if [[ "$(uname)" == Darwin ]]; then
         avfilter=libavfilter.10
         swscale=libswscale.8
         swresample=libswresample.5
+    elif [[ ${major_ver} == 8 ]]; then
+        avutil=libavutil.60
+        avcodec=libavcodec.62
+        avformat=libavformat.62
+        avdevice=libavdevice.62
+        avfilter=libavfilter.11
+        swscale=libswscale.9
+        swresample=libswresample.6
     else
         printf "Error: unexpected FFmpeg major version: %s\n"  ${major_ver}
         exit 1;

src/torchcodec/_core/BetaCudaDeviceInterface.cpp

@@ -203,10 +203,7 @@ BetaCudaDeviceInterface::BetaCudaDeviceInterface(const torch::Device& device)
   TORCH_CHECK(
       device_.type() == torch::kCUDA, "Unsupported device: ", device_.str());
 
-  // Initialize CUDA context with a dummy tensor
-  torch::Tensor dummyTensorForCudaInitialization = torch::empty(
-      {1}, torch::TensorOptions().dtype(torch::kUInt8).device(device_));
-
+  initializeCudaContextWithPytorch(device_);
   nppCtx_ = getNppStreamContext(device_);
 }
 

src/torchcodec/_core/CUDACommon.cpp

@@ -23,6 +23,14 @@ PerGpuCache<NppStreamContext> g_cached_npp_ctxs(
 
 } // namespace
 
+void initializeCudaContextWithPytorch(const torch::Device& device) {
+  // It is important for pytorch itself to create the cuda context. If ffmpeg
+  // creates the context it may not be compatible with pytorch.
+  // This is a dummy tensor to initialize the cuda context.
+  torch::Tensor dummyTensorForCudaInitialization = torch::zeros(
+      {1}, torch::TensorOptions().dtype(torch::kUInt8).device(device));
+}
+
 /* clang-format off */
 // Note: [YUV -> RGB Color Conversion, color space and color range]
 //

src/torchcodec/_core/CUDACommon.h

@@ -22,6 +22,8 @@ extern "C" {
 
 namespace facebook::torchcodec {
 
+void initializeCudaContextWithPytorch(const torch::Device& device);
+
 // Unique pointer type for NPP stream context
 using UniqueNppContext = std::unique_ptr<NppStreamContext>;
 

src/torchcodec/_core/CudaDeviceInterface.cpp

@@ -129,11 +129,10 @@ CudaDeviceInterface::CudaDeviceInterface(const torch::Device& device)
   TORCH_CHECK(
       device_.type() == torch::kCUDA, "Unsupported device: ", device_.str());
 
-  // It is important for pytorch itself to create the cuda context. If ffmpeg
-  // creates the context it may not be compatible with pytorch.
-  // This is a dummy tensor to initialize the cuda context.
-  torch::Tensor dummyTensorForCudaInitialization = torch::empty(
-      {1}, torch::TensorOptions().dtype(torch::kUInt8).device(device_));
+  initializeCudaContextWithPytorch(device_);
+
+  // TODO rename this, this is a hardware device context, not a CUDA context!
+  // See https://github.com/meta-pytorch/torchcodec/issues/924
   ctx_ = getCudaContext(device_);
   nppCtx_ = getNppStreamContext(device_);
 }

src/torchcodec/_core/DeviceInterface.h

@@ -32,8 +32,8 @@ struct DeviceInterfaceKey {
 
   explicit DeviceInterfaceKey(torch::DeviceType type) : deviceType(type) {}
 
-  DeviceInterfaceKey(torch::DeviceType type, const std::string_view& var)
-      : deviceType(type), variant(var) {}
+  DeviceInterfaceKey(torch::DeviceType type, const std::string_view& variant)
+      : deviceType(type), variant(variant) {}
 };
 
 class DeviceInterface {

test/test_decoders.py

@@ -591,7 +591,7 @@ def test_get_frame_at_av1(self, device):
         if device == "cuda" and get_ffmpeg_major_version() == 4:
             return
 
-        if device == "cuda" and in_fbcode():
+        if "cuda" in device and in_fbcode():
             pytest.skip("decoding on CUDA is not supported internally")
 
         decoder = VideoDecoder(AV1_VIDEO.path, device=device)
@@ -1477,6 +1477,9 @@ def test_beta_cuda_interface_get_frame_at(
         if seek_mode == "approximate" and not supports_approximate_mode(asset):
             pytest.skip("asset doesn't work with approximate mode")
 
+        if in_fbcode() and asset is AV1_VIDEO:
+            pytest.skip("AV1 CUDA not supported internally")
+
         ref_decoder = VideoDecoder(asset.path, device="cuda", seek_mode=seek_mode)
         beta_decoder = VideoDecoder(
             asset.path, device="cuda:0:beta", seek_mode=seek_mode
@@ -1522,6 +1525,8 @@ def test_beta_cuda_interface_get_frames_at(
     ):
         if seek_mode == "approximate" and not supports_approximate_mode(asset):
             pytest.skip("asset doesn't work with approximate mode")
+        if in_fbcode() and asset is AV1_VIDEO:
+            pytest.skip("AV1 CUDA not supported internally")
 
         ref_decoder = VideoDecoder(asset.path, device="cuda", seek_mode=seek_mode)
         beta_decoder = VideoDecoder(
@@ -1566,6 +1571,8 @@ def test_beta_cuda_interface_get_frames_at(
     def test_beta_cuda_interface_get_frame_played_at(self, asset, seek_mode):
         if seek_mode == "approximate" and not supports_approximate_mode(asset):
             pytest.skip("asset doesn't work with approximate mode")
+        if in_fbcode() and asset is AV1_VIDEO:
+            pytest.skip("AV1 CUDA not supported internally")
 
         ref_decoder = VideoDecoder(asset.path, device="cuda", seek_mode=seek_mode)
         beta_decoder = VideoDecoder(
@@ -1607,6 +1614,8 @@ def test_beta_cuda_interface_get_frame_played_at(self, asset, seek_mode):
     def test_beta_cuda_interface_get_frames_played_at(self, asset, seek_mode):
         if seek_mode == "approximate" and not supports_approximate_mode(asset):
             pytest.skip("asset doesn't work with approximate mode")
+        if in_fbcode() and asset is AV1_VIDEO:
+            pytest.skip("AV1 CUDA not supported internally")
 
         ref_decoder = VideoDecoder(asset.path, device="cuda", seek_mode=seek_mode)
         beta_decoder = VideoDecoder(
@@ -1649,6 +1658,8 @@ def test_beta_cuda_interface_get_frames_played_at(self, asset, seek_mode):
     def test_beta_cuda_interface_backwards(self, asset, seek_mode):
         if seek_mode == "approximate" and not supports_approximate_mode(asset):
             pytest.skip("asset doesn't work with approximate mode")
+        if in_fbcode() and asset is AV1_VIDEO:
+            pytest.skip("AV1 CUDA not supported internally")
 
         ref_decoder = VideoDecoder(asset.path, device="cuda", seek_mode=seek_mode)
         beta_decoder = VideoDecoder(