biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[cpu][perf] Accelerate unquantized-linear for AArch64 through oneDNN/ACL and weight prepack (#25948)

Browse Source 
Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
Co-authored-by: Li, Jiang <jiang1.li@intel.com>
This commit is contained in:
Fadi Arafeh
2025-10-04 05:16:38 +01:00
committed by GitHub
parent 2f7dbc9b42
commit 9705fba7b7
8 changed files with 111 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
csrc/cpu/dnnl_kernels.cpp
View File

@@ -527,21 +527,42 @@ void onednn_mm(torch::Tensor& c, // [M, OC], row-major
MatMulPrimitiveHandler* ptr =
reinterpret_cast<MatMulPrimitiveHandler*>(handler);
// ACL matmuls expect contiguous source tensors
#ifdef VLLM_USE_ACL
torch::Tensor a_contig = a.contiguous();
#endif
MatMulPrimitiveHandler::ExecArgs exec_args;
#ifdef VLLM_USE_ACL
exec_args.a_m_size = a_contig.size(0);
exec_args.a_m_stride = a_contig.stride(0);
#else
exec_args.a_m_size = a.size(0);
exec_args.a_m_stride = a.stride(0);
#endif
VLLM_DISPATCH_FLOATING_TYPES(a.scalar_type(), "onednn_mm", [&] {
if (bias.has_value()) {
exec_args.use_bias = true;
exec_args.bias_type = get_dnnl_type<scalar_t>();
#ifdef VLLM_USE_ACL
// ACL matmuls in oneDNN do not support a bias.
// We handle a matmul with bias by doing: c = bias; c += matmul(a, b)
c.copy_(bias.value());
#else
exec_args.bias_ptr = bias->data_ptr<scalar_t>();
#endif
} else {
exec_args.use_bias = false;
exec_args.bias_type = get_dnnl_type<void>();
exec_args.bias_ptr = nullptr;
}
#ifdef VLLM_USE_ACL
exec_args.a_ptr = a_contig.data_ptr<scalar_t>();
#else
exec_args.a_ptr = a.data_ptr<scalar_t>();
#endif
exec_args.c_ptr = c.data_ptr<scalar_t>();
ptr->execute(exec_args);