[CPU][Perf] Accelerate unquantized MoE for AArch64 (#46353)

Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
This commit is contained in:
Fadi Arafeh
2026-06-24 14:14:35 +00:00
committed by GitHub
parent 93ec645878
commit 061043eaca
15 changed files with 672 additions and 32 deletions
+2 -2
View File
@@ -538,7 +538,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
#endif
// fused moe
#if defined(__AVX512F__)
#if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT))
ops.def(
"prepack_moe_weight(Tensor weight, Tensor(a1!) packed_weight, str isa) "
"-> ()");
@@ -549,7 +549,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
"bool skip_weighted, "
"str act, str isa) -> ()");
ops.impl("cpu_fused_moe", torch::kCPU, &cpu_fused_moe);
#endif
#endif // #if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT))
ops.def(
"mla_decode_kvcache("
" Tensor! out, Tensor query, Tensor kv_cache,"