Enable u8i8 and bf16 MHA tokenization with transpose_b=true

v-Golubev · Nov 19, 2024 · 1de39e8 · 1de39e8
1 parent 0fb5b64
commit 1de39e8
Showing 1 changed file with 6 additions and 3 deletions.
diff --git a/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp b/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp
@@ -1018,9 +1018,12 @@ void Transformations::MainSnippets(void) {
         // Only FP32 dynamic MHA is supported
         if (matmul->is_dynamic())
             return false;
-        // [114487] brgemm kernel in oneDNN requires brgemm_copy_b kernel if MatMul node has transposed_b=True
-        // The current solution with ExtractExplicitMatMulTranspose pass is slower for non-f32 cases than using of brgemm_copy_b kernel
-        if (matmul->get_transpose_a() || matmul->get_transpose_b())
+        // Ticket 157340: repacking extraction is not supported for i8i8 case.
+        // If the repacking is performed inside the kernel, it may lead to performance degradation.
+        if (is_int8 && matmul->get_transpose_b())
+            return false;
+
+        if (matmul->get_transpose_a())
             return false;
         // [150842] The execution of Brgemm INT8/BF16 on AMX platforms depends on the value of "K % VNNIFactor".
         //          For more details, please teake a look at the ticket 150842