diff --git a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq.h b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq.h
index 91b59336b8c134566c7fec8dd4f1ea8b957aaa97..98efc8776260ea896e2a492f32bbb9b088003f44 100644
--- a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq.h
+++ b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq.h
@@ -214,7 +214,7 @@ void GmmDeq(
 CATLASS_DEVICE
 void BarrierBetweenUpAndDown()
 {
-    AscendC::PipeBarrier<PIPE_ALL>();
+    // AscendC::PipeBarrier<PIPE_ALL>();
     Arch::CrossCoreFlag gmm1AivFinished{0};
     if constexpr (g_coreType == AscendC::AIV) {
         Arch::CrossCoreBarrier<0x0, PIPE_MTE3>();
diff --git a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n128.cpp b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n128.cpp
index c7a48744f53ca7cdd6c866d33c247d955a41d7a2..08a351f5f4755bb2fc4c161b4e289061f2d09a4c 100644
--- a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n128.cpp
+++ b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n128.cpp
@@ -49,7 +49,7 @@ void gmm_deq_swiglu_quant_gmm_deq_n128(
     workspaceOffset += RoundUp<GM_ALIGN_BYTE>(static_cast<size_t>(m) * nOut * sizeof(int8_t));
     GM_ADDR gmPerTokenScale2 = workspace + workspaceOffset;
     workspaceOffset += RoundUp<GM_ALIGN_BYTE>(static_cast<size_t>(m) * sizeof(float));
-    GM_ADDR gmWorkspace = workspace + workspaceOffset;
+    GM_ADDR gmWorkspace = workspace + workspaceOffset - 512;
 
     using Gmm1L1TileShape = GemmShape<Gmm1TileArgs<PERMUTE_N128>::L1M, Gmm1TileArgs<PERMUTE_N128>::L1N, GMM1_L1K>;
     using Gmm1L0TileShape = GemmShape<Gmm1L1TileShape::M, Gmm1L1TileShape::N, GMM1_L0K>;
diff --git a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n256.cpp b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n256.cpp
index 32a5ce645ba9fb593c3fb75b5c8829ba2528bcbb..b689eec2b836db4f6f68c5a7ea739a433127b22c 100644
--- a/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n256.cpp
+++ b/src/kernels/mixkernels/gmm_deq_swiglu_quant_gmm_deq/op_kernel/gmm_deq_swiglu_quant_gmm_deq_n256.cpp
@@ -49,7 +49,7 @@ void gmm_deq_swiglu_quant_gmm_deq_n256(
     workspaceOffset += RoundUp<GM_ALIGN_BYTE>(static_cast<size_t>(m) * nOut * sizeof(int8_t));
     GM_ADDR gmPerTokenScale2 = workspace + workspaceOffset;
     workspaceOffset += RoundUp<GM_ALIGN_BYTE>(static_cast<size_t>(m) * sizeof(float));
-    GM_ADDR gmWorkspace = workspace + workspaceOffset;
+    GM_ADDR gmWorkspace = workspace + workspaceOffset - 512;
 
     using Gmm1L1TileShape = GemmShape<Gmm1TileArgs<PERMUTE_N256>::L1M, Gmm1TileArgs<PERMUTE_N256>::L1N, GMM1_L1K>;
     using Gmm1L0TileShape = GemmShape<Gmm1L1TileShape::M, Gmm1L1TileShape::N, GMM1_L0K>;