This issue collects tasks that block porting forced_align/cpu/compute.cpp and forced_align/gpu/compute.cu to use torch stable ABI.

• expose AT_DISPATCH_FLOATING_TYPES_AND_HALF to stable ABI, currently one need to implement the dispatch logic using switch block. Not a blocker but would be nice to have.
  For a workaround example, see
  
  
  

  audio/src/libtorchaudio/forced_align/cpu/compute.cpp

  Lines 184 to 221 in b3a5a0e

  	switch (logProbs.scalar_type()) {
  	case ScalarType::Double: {
  	if (targets.scalar_type() == ScalarType::Long) {
  	forced_align_impl<double, ScalarType::Long>(logProbs, targets, blank, paths);
  	} else if (targets.scalar_type() == ScalarType::Int) {
  	forced_align_impl<double, ScalarType::Int>(logProbs, targets, blank, paths);
  	} else {
  	STD_TORCH_CHECK(false, "unreachable");
  	}
  	break;
  	}
  	case ScalarType::Float: {
  	if (targets.scalar_type() == ScalarType::Long) {
  	forced_align_impl<float, ScalarType::Long>(logProbs, targets, blank, paths);
  	} else if (targets.scalar_type() == ScalarType::Int) {
  	forced_align_impl<float, ScalarType::Int>(logProbs, targets, blank, paths);
  	} else {
  	STD_TORCH_CHECK(false, "unreachable");
  	}
  	break;
  	}
  	case ScalarType::Half: {
  	if (targets.scalar_type() == ScalarType::Long) {
  	forced_align_impl<c10::Half, ScalarType::Long>(logProbs, targets, blank, paths);
  	} else if (targets.scalar_type() == ScalarType::Int) {
  	forced_align_impl<c10::Half, ScalarType::Int>(logProbs, targets, blank, paths);
  	} else {
  	STD_TORCH_CHECK(false, "unreachable");
  	}
  	break;
  	}
  	default: {
  	STD_TORCH_CHECK(false, "unreachable");
  	}
  	};
  	return std::make_tuple(paths, logProbs);
  	}

  
  
• implement item<T>() as a torch::stable::Tensor template method. A workaround is implemented in
  
  
  

  audio/src/libtorchaudio/utils.h

  Lines 19 to 33 in b3a5a0e

  	template <typename T>
  	T item(const torch::stable::Tensor& t) {
  	STD_TORCH_CHECK(t.numel() == 1, "item requires single element tensor input");
  	if (t.is_cpu()) {
  	return t.const_data_ptr<T>()[0];
  	#ifdef USE_CUDA
  	} else if (t.is_cuda()) {
  	T value;
  	C10_CUDA_CHECK(cudaMemcpyAsync(&value, t.data_ptr(), sizeof(T), cudaMemcpyDeviceToHost));
  	return value;
  	#endif
  	} else {
  	STD_TORCH_CHECK(false, "unreachable");
  	}
  	}

  
  
• implement accessor template as a torch::stable::Tensor template method or replace its usage. Fix available: [STABLE ABI] Add accessor template method to torch::stable::Tensor pytorch#161967
• expose PackedTensorAccessor32 to stable ABI or replace its usage
  Fix available: [STABLE ABI] Add packed_accessor32 and generic_packed_accessor template methods to torch::stable::Tensor pytorch#161897
• implement to as a torch::stable::Tensor method
  Workarounds: use cpu() op from [STABLE ABI] Add cpu operation. pytorch#161911 or new_empty/copy_ approach as in
  
  
  

  audio/src/libtorchaudio/forced_align/gpu/compute.cu

  Lines 335 to 340 in b3a5a0e

  	Tensor pathsCuda = torch::stable::new_empty(paths,
  	torchaudio::util::sizes(paths),
  	std::nullopt,
  	aoti_torch_device_type_cuda(),
  	logProbs.get_device_index());
  	torch::stable::copy_(pathsCuda, paths);

  
  

^ @NicolasHug @scotts @janeyx99