[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
+4 -3
View File
@@ -162,12 +162,13 @@ 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)
#endif
// CPU tensor -> CUDA UVA view (shared CUDA/ROCm)
// TODO: Move this out once ROCm upgrade their torch to 2.11.
// CPU tensor -> CUDA UVA view (shared CUDA)
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,