[CPU] Migrate unquantized MoE to the modular-kernel experts structure (#50133)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Fadi Arafeh <fadi.arafeh@arm.com>
This commit is contained in:
Li, Jiang
2026-08-03 14:55:04 +08:00
committed by GitHub
co-authored by Claude Opus 5 Fadi Arafeh
parent b3f97dae24
commit 0a6446005d
20 changed files with 972 additions and 839 deletions
-3
View File
@@ -614,7 +614,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
#endif
// fused moe
#if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT) && !defined(__APPLE__))
ops.def(
"prepack_moe_weight(Tensor weight, Tensor(a1!) packed_weight, str isa) "
"-> ()");
@@ -625,8 +624,6 @@ 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 // #if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT) &&
// !defined(__APPLE__))
#if defined(ARM_I8MM_SUPPORT) && defined(ARM_BF16_SUPPORT) && \
!defined(__APPLE__)
ops.def(