[ROCm] [PyTorch] Move to stable abi since ROCm upgraded to torch 2.11 (#47128)

Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
TJian
2026-07-02 18:34:07 +08:00
committed by GitHub
parent 84b9c2762f
commit de2a8fc042
6 changed files with 22 additions and 131 deletions
+3 -4
View File
@@ -181,13 +181,12 @@ torch::stable::Tensor awq_dequantize(torch::stable::Tensor _kernel,
// AllSpark ops: declarations are in the source files
// (allspark_repack.cu and allspark_qgemm_w8a16.cu)
// TODO: Move this out once ROCm upgrade their torch to 2.11.
// CPU tensor -> CUDA UVA view (shared CUDA)
#endif
// CPU tensor -> CUDA UVA view (shared CUDA/ROCm)
torch::stable::Tensor get_cuda_view_from_cpu_tensor(
torch::stable::Tensor& cpu_tensor);
#endif
// Attention kernels (shared CUDA/ROCm)
void merge_attn_states(
torch::stable::Tensor& output,
+2 -7
View File
@@ -29,11 +29,10 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
"()");
ops.def("permute_cols(Tensor A, Tensor perm) -> Tensor");
#ifndef USE_ROCM
// TODO: Remove this once ROCm upgrade to torch 2.11.
ops.def("get_cuda_view_from_cpu_tensor(Tensor cpu_tensor) -> Tensor");
#ifndef USE_ROCM
// Note about marlin kernel 'workspace' arguments:
// Technically these should be mutable since they are modified by the kernel.
// But since they are set back to zero once the kernel is finished we can
@@ -753,8 +752,6 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
ops.impl("paged_attention_v2", TORCH_BOX(&paged_attention_v2));
}
// TODO: Remove this once ROCm upgrade to torch 2.11.
#ifndef USE_ROCM
STABLE_TORCH_LIBRARY_IMPL(_C, CPU, ops) {
ops.impl("get_cuda_view_from_cpu_tensor",
TORCH_BOX(&get_cuda_view_from_cpu_tensor));
@@ -773,8 +770,6 @@ STABLE_TORCH_LIBRARY_IMPL(_C_cuda_utils, CompositeExplicitAutograd,
TORCH_BOX(&get_max_shared_memory_per_block_device_attribute));
}
#endif
// These capability-check functions take only primitive args (no tensors), so
// there is no device to dispatch on. CompositeExplicitAutograd makes them
// available for all backends. This is the stable ABI equivalent of calling