Port C++ code to rely on torch stable ABI API. #1188

NicolasHug · 2026-01-23T11:42:15Z

No description provided.

NicolasHug · 2026-01-23T14:47:21Z

test/third-party-interface/ThirdPartyInterfaceTest.cpp

 class DummyDeviceInterface : public DeviceInterface {
 public:
-  DummyDeviceInterface(const torch::Device& device) : DeviceInterface(device) {}
+  DummyDeviceInterface(const StableDevice& device) : DeviceInterface(device) {}


@dvrogozh , heads up that this is something we might want to do in the near future.

The general context is that we want to rely on torch stable ABI so that we (or you) don't have to synchronize torchcodec releases with torch releases 6X a year.

I think it means that the third-party extensions will have to pass-in "Stable" types - you might still be able to use the non-stable ones internally as long as you convert them.
Let me know if you have any concern about that.

Note that the PR is in a very, very WIP state at the moment. Plenty of things are likely to change. But this is the direction we're heading to.

@NicolasHug , thank your for heads up. I will start looking into the story, but "stable" sounds very appealing. Once PR will be in a closer to merge stage I would love to give it a try on my side (create respective PR for xpu and see how it goes).

NicolasHug · 2026-01-23T14:47:50Z

src/torchcodec/share/cmake/TorchCodec/TorchCodecConfig.cmake

+      # Set PyTorch stable ABI target version to 2.11 for consumers.
+      # This ensures third-party code uses stable ABI types compatible with
+      # the library. Format: 0x MAJOR(2) MINOR(2) PATCH(2) TWEAK(2) zeros(8)
+      INTERFACE_COMPILE_DEFINITIONS "TORCH_TARGET_VERSION=0x020b000000000000"


TODO: I'm not sure we'll want to enforce that on third-parties?

CC @dvrogozh

NicolasHug · 2026-01-23T14:53:09Z

src/torchcodec/_core/__init__.py

 )
 from .ops import (
    _add_video_stream,
+    _destroy_decoder,


TODO: there's no way we want that. Aparently this is because from_blob doesn't support deleters? See comment in custom_ops.cpp.

NicolasHug · 2026-01-23T16:32:33Z

src/torchcodec/_core/AVIOTensorContext.h

 // For Decoding: enables users to pass in the entire video or audio as bytes.
 // Our read and seek functions then traverse the bytes in memory.
-class AVIOFromTensorContext : public AVIOContextHolder {
+class TORCHCODEC_API AVIOFromTensorContext : public AVIOContextHolder {


TODO: understand why that's needed. Ideally it shouldn't.

NicolasHug · 2026-01-23T16:32:52Z

src/torchcodec/_core/BetaCudaDeviceInterface.cpp

+// Enable CUDA-specific functions in PyTorch stable ABI headers
+#define USE_CUDA 1


NicolasHug · 2026-01-23T16:33:15Z

src/torchcodec/_core/BetaCudaDeviceInterface.cpp

+  if (!device_.has_index()) {
+    int resolvedIndex = getDeviceIndex(device_);
+    device_.set_index(static_cast<StableDeviceIndex>(resolvedIndex));
+  }


TODO above: this was causing a hang - investigate.

NicolasHug · 2026-01-23T16:35:56Z

src/torchcodec/_core/CMakeLists.txt

+    target_include_directories(${library_name}
+        BEFORE PRIVATE
+        ${pybind11_INCLUDE_DIRS}
+    )


TODO investigate above.

NicolasHug · 2026-01-23T16:36:25Z

src/torchcodec/_core/CMakeLists.txt

+            "-Wno-attributes"
+        )
+    endif()
+


TODO on all these

NicolasHug · 2026-01-23T16:37:04Z

src/torchcodec/_core/CpuDeviceInterface.cpp

+    // For a tensor of shape [numChannels, numSamples], we need to calculate
+    // offsets
+    outputBuffers[i] = static_cast<uint8_t*>(lastSamples.mutable_data_ptr()) +
+        i * numRemainingSamples * sizeof(float);


NicolasHug · 2026-01-23T16:38:03Z

src/torchcodec/_core/custom_ops.cpp

+  // Note: impl_abstract_pystub is not available in stable ABI.
+  // The Python stubs are handled via Python-side registration instead.


NicolasHug · 2026-01-23T16:38:23Z

src/torchcodec/_core/custom_ops.cpp

+// exits. Note: This map is not thread-safe; concurrent decoder
+// creation/destruction from multiple threads is not supported.
+std::unordered_map<SingleStreamDecoder*, std::unique_ptr<SingleStreamDecoder>>
+    g_decoder_registry;


TODO: we reaaaally want to avoid the new logic. Can we add deleter support to the stable ABI from_blob?

NicolasHug · 2026-01-23T16:38:37Z

src/torchcodec/_core/custom_ops.cpp

+      "_add_video_stream(Tensor(a!) decoder, *, int? num_threads=None, str? dimension_order=None, int? stream_index=None, str device=\"cpu\", str device_variant=\"ffmpeg\", str transform_specs=\"\", Tensor? custom_frame_mappings_all_frames=None, Tensor? custom_frame_mappings_is_key_frame=None, Tensor? custom_frame_mappings_duration=None, str? color_conversion_library=None) -> ()");
  m.def(
-      "add_video_stream(Tensor(a!) decoder, *, int? num_threads=None, str? dimension_order=None, int? stream_index=None, str device=\"cpu\", str device_variant=\"ffmpeg\", str transform_specs=\"\", (Tensor, Tensor, Tensor)? custom_frame_mappings=None) -> ()");
+      "add_video_stream(Tensor(a!) decoder, *, int? num_threads=None, str? dimension_order=None, int? stream_index=None, str device=\"cpu\", str device_variant=\"ffmpeg\", str transform_specs=\"\", Tensor? custom_frame_mappings_all_frames=None, Tensor? custom_frame_mappings_is_key_frame=None, Tensor? custom_frame_mappings_duration=None) -> ()");


TODO: can't pass tuples anymore?

NicolasHug · 2026-01-23T16:39:15Z

src/torchcodec/_core/custom_ops.cpp

+    StableTensor& decoder,
    std::optional<int64_t> num_threads = std::nullopt,
-    std::optional<std::string_view> dimension_order = std::nullopt,
+    std::optional<std::string> dimension_order = std::nullopt,


TODO we went from string_view to string everywhere, is that OK? Is that needed?

NicolasHug · 2026-01-23T16:40:46Z

src/torchcodec/_core/DeviceInterface.cpp

+
+  // Allocate output tensor
+  // Note: Stable ABI's from_blob doesn't support custom deleters,
+  // so we allocate and copy the data instead


NicolasHug · 2026-01-23T16:41:49Z

src/torchcodec/_core/ops.py

+# Register atexit handler to clear all decoders before Python shuts down.
+# This prevents hangs when decoders holding file-like objects are destroyed
+# during static destruction (after Python has started shutting down).
+atexit.register(_clear_all_decoders)


NicolasHug added 7 commits January 22, 2026 17:58

WIP

5025986

It compiles. And crashes.

68db6f2

Compiles, import is OK, segfaults on tests

04b3e5c

Some tests are passing

f1bb56f

Really weird fix to fix hanging issue

ace2198

More stufffffff

0a5dc8d

Lint

0b7b31a

meta-cla bot added the CLA Signed This label is managed by the Meta Open Source bot. label Jan 23, 2026

NicolasHug added 6 commits January 23, 2026 13:18

Fix stuff?

3f0219a

lint

22d811a

more

8e10fab

refact C++ test

b0e00b7

Fix Windows?

b86c98b

lint

72977df

NicolasHug commented Jan 23, 2026

View reviewed changes

src/torchcodec/_core/CMakeLists.txt

"-Wno-attributes"

)

endif()

Copy link

Contributor Author

NicolasHug Jan 23, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

TODO on all these

NicolasHug commented Jan 23, 2026

View reviewed changes

NicolasHug mentioned this pull request Jan 26, 2026

Add deleter support to torch::stable::from_blob pytorch/pytorch#173371

Open

		// Enable CUDA-specific functions in PyTorch stable ABI headers
		#define USE_CUDA 1

		// Note: impl_abstract_pystub is not available in stable ABI.
		// The Python stubs are handled via Python-side registration instead.

Port C++ code to rely on torch stable ABI API. #1188

Are you sure you want to change the base?

Port C++ code to rely on torch stable ABI API. #1188

Uh oh!

Conversation

NicolasHug commented Jan 23, 2026

Uh oh!

NicolasHug Jan 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

NicolasHug Jan 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

NicolasHug Jan 23, 2026 •

edited

Loading

NicolasHug Jan 23, 2026 •

edited

Loading