[Bugfix] [ROCm] [Critical] fallback to regular abi for ROCm (#44648)

Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
TJian
2026-06-05 15:51:17 +00:00
committed by GitHub
parent 6a894574bf
commit aa6fb8a329
6 changed files with 127 additions and 25 deletions
+20 -14
View File
@@ -28,9 +28,11 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
"output_s, int group_size, float eps, float int8_min, float int8_max) -> "
"()");
#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
ops.def("permute_cols(Tensor A, Tensor perm) -> Tensor");
#endif
@@ -676,11 +678,28 @@ 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));
}
STABLE_TORCH_LIBRARY_FRAGMENT(_C_cuda_utils, cuda_utils) {
cuda_utils.def("get_device_attribute(int attribute, int device_id) -> int");
cuda_utils.def(
"get_max_shared_memory_per_block_device_attribute(int device_id) -> int");
}
STABLE_TORCH_LIBRARY_IMPL(_C_cuda_utils, CompositeExplicitAutograd,
cuda_utils) {
cuda_utils.impl("get_device_attribute", TORCH_BOX(&get_device_attribute));
cuda_utils.impl("get_max_shared_memory_per_block_device_attribute",
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
@@ -701,19 +720,6 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CompositeExplicitAutograd, ops) {
ops.impl("ggml_moe_get_block_size", TORCH_BOX(&ggml_moe_get_block_size));
}
STABLE_TORCH_LIBRARY_FRAGMENT(_C_cuda_utils, cuda_utils) {
cuda_utils.def("get_device_attribute(int attribute, int device_id) -> int");
cuda_utils.def(
"get_max_shared_memory_per_block_device_attribute(int device_id) -> int");
}
STABLE_TORCH_LIBRARY_IMPL(_C_cuda_utils, CompositeExplicitAutograd,
cuda_utils) {
cuda_utils.impl("get_device_attribute", TORCH_BOX(&get_device_attribute));
cuda_utils.impl("get_max_shared_memory_per_block_device_attribute",
TORCH_BOX(&get_max_shared_memory_per_block_device_attribute));
}
// Cache ops
STABLE_TORCH_LIBRARY_FRAGMENT(_C_cache_ops, ops) {
// Swap in (out) the cache blocks from src to dst.