diff --git a/tests/unit/test_pv_only.cu b/tests/unit/test_pv_only.cu
index 9c5bcbfb..74c732c4 100644
--- a/tests/unit/test_pv_only.cu
+++ b/tests/unit/test_pv_only.cu
@@ -47,10 +47,11 @@ test_pv_only_kernel(
     size_t off = 0;
     uint32_t* sTmemBase = (uint32_t*)(sbuf + off); off = 4;
     off = (off + 15) & ~(size_t)15;
+    float* s_p_vals = (float*)(sbuf + off); off += SK * sizeof(float);
+    off = (off + 15) & ~(size_t)15;
     bf16_t* sPk = (bf16_t*)(sbuf + off); off += TILE_SZ * sizeof(bf16_t);
     off = (off + 127) & ~(size_t)127;
     bf16_t* sV = (bf16_t*)(sbuf + off); off += V_SUB_SZ * sizeof(bf16_t);
-    float* s_p_vals = (float*)(sbuf + off); off += SK * sizeof(float);
 
     // Load P values into SMEM
     for (int j = tid; j < SK; j += 128) s_p_vals[j] = p_input[j];
@@ -112,7 +113,7 @@ test_pv_only_kernel(
 }
 
 int main() {
-    printf("PV-Only Test (HD=%d, SK=%d)\n", HD);
+    printf("PV-Only Test (HD=%d, SK=%d)\n", (int)HD, (int)SK);
 
     bf16_t* h_v = (bf16_t*)malloc(HD * SK * sizeof(bf16_t));
     float* h_p = (float*)malloc(SK * sizeof(float));
@@ -128,7 +129,7 @@ int main() {
     cudaMemcpy(d_v, h_v, HD * SK * sizeof(bf16_t), cudaMemcpyHostToDevice);
     cudaMemcpy(d_p, h_p, SK * sizeof(float), cudaMemcpyHostToDevice);
 
-    int smem = 4 + 16 + TILE_SZ + V_SUB_SZ*2 + SK*4 + 256;
+    size_t smem = 4 + 16 + SK*4 + 16 + TILE_SZ*2 + 128 + V_SUB_SZ*2 + 256;
     cudaFuncSetAttribute(test_pv_only_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem);
     test_pv_only_kernel<<<1, 128, smem>>>(d_out, d_v, d_p);
     cudaError_t err = cudaDeviceSynchronize();