diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 50e18f57a22f7..d73ee6e0afd5e 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -343,13 +343,13 @@ extern "C" {
     GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
     GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
 
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) ||  defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
     enum ggml_compute_backend_type {
         GGML_COMPUTE_BACKEND_CPU=0,
         GGML_COMPUTE_BACKEND_TSAVORITE,
         GGML_COMPUTE_BACKEND_COUNT
     };
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GML_PERF-related flag */
 
     enum ggml_status {
         GGML_STATUS_ALLOC_FAILED = -2,
@@ -660,14 +660,14 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
         int64_t perf_runs;
         int64_t perf_time_us;
         enum ggml_compute_backend_type ggml_compute_backend;
         char padding[4];
 #else
         char padding[8];
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GML_PERF-related flag */
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -2557,7 +2557,7 @@ extern "C" {
     GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
     GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
  
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
 struct ggml_perf_backend_subtotals {
     int64_t total_us;
     int64_t runs;
@@ -2587,7 +2587,7 @@ void ggml_perf_write_detailed_csv(struct ggml_cgraph * cgraph, FILE *fp);
 void ggml_perf_accumulate(struct ggml_perf_totals totals[GGML_OP_COUNT], struct ggml_cgraph * cgraph);
 const char * ggml_backend_type(enum ggml_compute_backend_type backend);
 
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GML_PERF-related flags */
 
 #ifdef  __cplusplus
 }
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index 3654fc3185ae6..527288c6239db 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -2879,12 +2879,12 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
     for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
 
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
         int64_t t_start = ggml_time_us();
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GGML_PERF-related flags */
         ggml_compute_forward(&params, node);
 
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
         int64_t t_end = ggml_time_us();
         node->perf_runs++;
         if (t_end >= t_start) {
@@ -2893,7 +2893,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
             // Handle wraparound by assuming timer rolls over at max int64_t value
             node->perf_time_us += (INT64_MAX - t_start + t_end + 1);
         }
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GGML_PERF-related flags */
         if (state->ith == 0 && cplan->abort_callback &&
                 cplan->abort_callback(cplan->abort_callback_data)) {
             atomic_store_explicit(&tp->abort, node_n + 1, memory_order_relaxed);
diff --git a/ggml/src/ggml-tsavorite/ggml-tsavorite.cpp b/ggml/src/ggml-tsavorite/ggml-tsavorite.cpp
index 30de81e576b67..430e1894c5015 100644
--- a/ggml/src/ggml-tsavorite/ggml-tsavorite.cpp
+++ b/ggml/src/ggml-tsavorite/ggml-tsavorite.cpp
@@ -929,9 +929,9 @@ static enum ggml_status ggml_tsavorite_graph_compute(ggml_backend_t backend,
 
   for (int i = 0; i < cgraph->n_nodes; i++) {
      int32_t kernel_sub_type=-1;
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
     int64_t t_start = ggml_time_us();
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GGML_PERF-related flags */
     node = cgraph->nodes[i];
     src0 = node->src[0];
     src1 = node->src[1];
@@ -1279,7 +1279,7 @@ static enum ggml_status ggml_tsavorite_graph_compute(ggml_backend_t backend,
           device->stats.op_run_count[kernel_type].max_num_of_elem < max_num_of_elem)
         device->stats.op_run_count[kernel_type].max_num_of_elem = max_num_of_elem;
     }
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
     int64_t t_end = ggml_time_us();
     node->perf_runs++;
     node->ggml_compute_backend = GGML_COMPUTE_BACKEND_TSAVORITE;
@@ -1289,7 +1289,7 @@ static enum ggml_status ggml_tsavorite_graph_compute(ggml_backend_t backend,
         // Handle wraparound by assuming timer rolls over at max int64_t value
         node->perf_time_us += (INT64_MAX - t_start + t_end + 1);
     }
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GGML_PERF-related flags */
   }
 
   // This this need to implement correctly when we have mixture of CPU and accelerator operation
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 6adf7849d80f2..960622c9cb8ba 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1020,12 +1020,12 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "GLU",
 };
 
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
 static const char * GGML_BACKEND_TYPE[GGML_COMPUTE_BACKEND_COUNT] = {
     "CPU",
     "OPU"
 };
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GGML_PERF-related flags */
 
 static_assert(GGML_OP_COUNT == 90, "GGML_OP_COUNT != 90");
 
@@ -1262,11 +1262,11 @@ const char * ggml_op_name(enum ggml_op op) {
     return GGML_OP_NAME[op];
 }
 
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) ||  defined(GGML_PERF_DETAIL)
 const char * ggml_backend_type(enum ggml_compute_backend_type backend) {
     return GGML_BACKEND_TYPE[backend];
 }
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GML_PERF-related flags */
 
 const char * ggml_op_symbol(enum ggml_op op) {
     return GGML_OP_SYMBOL[op];
@@ -1692,11 +1692,11 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.data         =*/ obj_alloc_size > 0 ? (void *)(result + 1) : data,
         /*.name         =*/ { 0 },
         /*.extra        =*/ NULL,
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) ||  defined(GGML_PERF_DETAIL)
         /*.perf_runs    =*/ 0,
         /*.perf_time_us =*/ 0,
         /*.ggml_compute_backend =*/ GGML_COMPUTE_BACKEND_CPU,
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GML_PERF-related flags */
         /*.padding      =*/ { 0 },
     };
 
@@ -7231,7 +7231,7 @@ bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, cons
     return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0;
 }
 
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) ||  defined(GGML_PERF_DETAIL)
 void ggml_perf_accumulate(struct ggml_perf_totals totals[GGML_OP_COUNT], struct ggml_cgraph * cgraph) {
     for (int i = 0; i < cgraph->n_nodes; ++i) {
         struct ggml_tensor * node = cgraph->nodes[i];
@@ -7258,7 +7258,7 @@ void ggml_perf_accumulate(struct ggml_perf_totals totals[GGML_OP_COUNT], struct
         }
     }
 }
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GML_PERF-related flags */
 
 #if defined(GGML_PERF_DETAIL)
 FILE * ggml_perf_log_open(const char *filename) {
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 37afd06db38ed..a7f77ae1c64f3 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -1090,12 +1090,12 @@ int llama_context::decode(const llama_batch & batch_inp) {
         ggml_status status;
         const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status);
 
-#if defined(GGML_PERF)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE)
         ggml_perf_accumulate(perf_totals, res->get_gf());
 #elif defined(GGML_PERF_DETAIL)
         ggml_perf_accumulate(perf_totals, res->get_gf());
         ggml_perf_write_detailed_csv(res->get_gf(), perf_all_shape_fp);
-#endif /* GGML_PERF || GGML_PERF_DETAIL */
+#endif /* GML_PERF-related flags */
 
 
         if (!res) {
@@ -2759,7 +2759,7 @@ llama_perf_context_data llama_perf_context(const llama_context * ctx) {
 }
 
 
-#if defined(GGML_PERF)
+#if defined(GGML_PERF_RELEASE)
 void ggml_perf_print_totals(struct ggml_perf_totals totals[GGML_OP_COUNT]) {
     LLAMA_LOG_TSAVORITE("\n=== GGML Perf Summary ===\n");
     LLAMA_LOG_TSAVORITE("  %-16s  %7s  %14s  %16s\n", "Op", "Runs", "Total us", "Avg us");
@@ -2788,7 +2788,7 @@ void ggml_perf_print_totals(struct ggml_perf_totals totals[GGML_OP_COUNT]) {
     }
 }
 
-#elif defined(GGML_PERF_DETAIL)
+#elif defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
 void ggml_perf_print_totals(struct ggml_perf_totals totals[GGML_OP_COUNT]) {
     LLAMA_LOG_TSAVORITE("\n=== GGML Perf Summary ===\n");
     LLAMA_LOG_TSAVORITE("  %-16s %-8s %7s  %14s  %16s\n", "Op", "Target", "Runs", "Total us", "Avg us");
@@ -2838,7 +2838,7 @@ void ggml_perf_print_totals(struct ggml_perf_totals totals[GGML_OP_COUNT]) {
         }
     }
 }
-#endif /* GGML_PERF  || GGML_PERF_DETAI */
+#endif /* GGML_PERF-related flags */
 
 
 void llama_perf_context_print(const llama_context * ctx) {
@@ -2852,7 +2852,7 @@ void llama_perf_context_print(const llama_context * ctx) {
             __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval);
     LLAMA_LOG_INFO("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
 
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
     LLAMA_LOG_TSAVORITE("\n%s:        load time = %10.2f ms\n", __func__, data.t_load_ms);
     LLAMA_LOG_TSAVORITE("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval);
@@ -2861,7 +2861,7 @@ void llama_perf_context_print(const llama_context * ctx) {
     LLAMA_LOG_TSAVORITE("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
 
     ggml_perf_print_totals(const_cast<ggml_perf_totals *>(ctx->perf_totals));
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GGML_PERF-related flags */
 }
 
 void llama_perf_context_reset(llama_context * ctx) {
diff --git a/src/llama-context.h b/src/llama-context.h
index e1a5bc4f9efdd..fe7c4cb579f23 100644
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -198,9 +198,9 @@ struct llama_context {
 
     // reserve a graph with a dummy ubatch of the specified size
     ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false);
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
     struct ggml_perf_totals perf_totals[GGML_OP_COUNT] = {};  // add this to llama_context
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GGML_PERF-related flags */
 
 private:
     llm_graph_params graph_params(
diff --git a/tools/main/main.cpp b/tools/main/main.cpp
index bd8c364584633..dc90b08a05296 100644
--- a/tools/main/main.cpp
+++ b/tools/main/main.cpp
@@ -126,9 +126,9 @@ int main(int argc, char ** argv) {
         LOG_WRN("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-#if defined(GGML_PERF) || defined(GGML_PERF_DETAIL)
+#if defined(GGML_PERF) || defined(GGML_PERF_RELEASE) || defined(GGML_PERF_DETAIL)
     llama_log_set(my_logger, nullptr);
-#endif /* GGML_PERF  || GGML_PERF_DETAIL */
+#endif /* GGML_PERF-related flags */
     LOG_INF("%s: llama backend init\n", __func__);
 
     llama_backend_init();
diff --git a/tsi-pkg-build.sh b/tsi-pkg-build.sh
index 6dcd6852b8e8d..fefb3b22f715e 100755
--- a/tsi-pkg-build.sh
+++ b/tsi-pkg-build.sh
@@ -32,10 +32,18 @@ cd ../posix-kernel/
 
 cd ../../
 
-#Compile for posix with build-posix as a target folder
+#Compile for posix & fpga with build-posix as a target folder
 
 echo 'building llama.cp, ggml for tsavorite  and other binary for posix'
-cmake -B build-posix -DGGML_TSAVORITE=ON -DGGML_TSAVORITE_TARGET=posix -DCMAKE_C_FLAGS="-DGGML_PERF"   -DCMAKE_CXX_FLAGS="-DGGML_PERF"
+if [ "$(echo "$1" | tr '[:upper:]' '[:lower:]')" = "release" ];
+then
+  cmake -B build-posix -DGGML_TSAVORITE=ON -DGGML_TSAVORITE_TARGET=posix -DCMAKE_C_FLAGS="-DGGML_PERF_RELEASE"   -DCMAKE_CXX_FLAGS="-DGGML_PERF_RELEASE"
+elif [ "$(echo "$1" | tr '[:upper:]' '[:lower:]')" = "debug" ]; then
+  cmake -B build-posix -DGGML_TSAVORITE=ON -DGGML_TSAVORITE_TARGET=posix -DCMAKE_C_FLAGS="-DGGML_PERF_DETAIL"   -DCMAKE_CXX_FLAGS="-DGGML_PERF_DETAIL"
+else
+  cmake -B build-posix -DGGML_TSAVORITE=ON -DGGML_TSAVORITE_TARGET=posix -DCMAKE_C_FLAGS="-DGGML_PERF"   -DCMAKE_CXX_FLAGS="-DGGML_PERF"
+fi
+
 cmake --build build-posix --config Release
 
 # Fix GLIBC compatibility for TSI binaries
@@ -64,12 +72,21 @@ chmod +x build-posix/bin/llama-cli
 echo 'building llama.cp, ggml for tsavorite  and other binary for fpga'
 export CC="/proj/rel/sw/arm-gnu-toolchain-14.2.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-none-linux-gnu-gcc"
 export CXX="/proj/rel/sw/arm-gnu-toolchain-14.2.rel1-x86_64-aarch64-none-linux-gnu/bin/aarch64-none-linux-gnu-g++"
-cmake -B build-fpga -DGGML_TSAVORITE=ON -DGGML_TSAVORITE_TARGET=fpga -DCMAKE_C_FLAGS="-DGGML_PERF"   -DCMAKE_CXX_FLAGS="-DGGML_PERF"
+
+if [ "$(echo "$1" | tr '[:upper:]' '[:lower:]')" = "release" ];
+then
+  cmake -B build-fpga -DGGML_TSAVORITE=ON -DGGML_TSAVORITE_TARGET=fpga -DCMAKE_C_FLAGS="-DGGML_PERF_RELEASE"   -DCMAKE_CXX_FLAGS="-DGGML_PERF_RELEASE"
+elif [ "$(echo "$1" | tr '[:upper:]' '[:lower:]')" = "debug" ]; then
+  cmake -B build-fpga -DGGML_TSAVORITE=ON -DGGML_TSAVORITE_TARGET=fpga -DCMAKE_C_FLAGS="-DGGML_PERF_DETAIL"   -DCMAKE_CXX_FLAGS="-DGGML_PERF_DETAIL"
+else
+  cmake -B build-fpga -DGGML_TSAVORITE=ON -DGGML_TSAVORITE_TARGET=fpga -DCMAKE_C_FLAGS="-DGGML_PERF"   -DCMAKE_CXX_FLAGS="-DGGML_PERF"
+fi
+
 cmake --build build-fpga --config Release
 
 
 echo 'creating tar bundle for fpga'
-TSI_GGML_VERSION=0.0.8
+TSI_GGML_VERSION=0.0.9
 TSI_GGML_BUNDLE_INSTALL_DIR=tsi-ggml
 GGML_TSI_INSTALL_DIR=ggml-tsi-kernel
 TSI_GGML_RELEASE_DIR=/proj/rel/sw/ggml