diff --git a/Makefile b/Makefile
index 3b48eec99..b297959c9 100644
--- a/Makefile
+++ b/Makefile
@@ -74,13 +74,17 @@ endif
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 	# Use all CPU extensions that are available:
-	CFLAGS += -march=native -mtune=native
+	CFLAGS   += -march=native -mtune=native
 	CXXFLAGS += -march=native -mtune=native
+
+	# Usage AVX-only
+	#CFLAGS   += -mfma -mf16c -mavx
+	#CXXFLAGS += -mfma -mf16c -mavx
 endif
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
-		CFLAGS += -mcpu=power9
+		CFLAGS   += -mcpu=power9
 		CXXFLAGS += -mcpu=power9
 	endif
 	# Require c++23's std::byteswap for big-endian support.
@@ -114,7 +118,7 @@ ifdef LLAMA_GPROF
 	CXXFLAGS += -pg
 endif
 ifneq ($(filter aarch64%,$(UNAME_M)),)
-	CFLAGS += -mcpu=native
+	CFLAGS   += -mcpu=native
 	CXXFLAGS += -mcpu=native
 endif
 ifneq ($(filter armv6%,$(UNAME_M)),)
diff --git a/ggml.c b/ggml.c
index 72b392fdb..46c0292fe 100644
--- a/ggml.c
+++ b/ggml.c
@@ -468,6 +468,14 @@ static inline int hsum_i32_8(const __m256i a) {
     return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
 }
 
+// horizontally add 4 int32_t
+static inline int hsum_i32_4(const __m128i a) {
+    const __m128i hi64 = _mm_unpackhi_epi64(a, a);
+    const __m128i sum64 = _mm_add_epi32(hi64, a);
+    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+
 #if __AVX2__ || __AVX512F__
 // Unpack 32 4-bit fields into 32 bytes
 // The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
@@ -1381,7 +1389,6 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
         y[i].s1 = d * sum1;
     }
 #elif defined(__AVX2__) || defined(__AVX__)
-    // TODO !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
     for (int i = 0; i < nb; i++) {
         // Load elements into 4 AVX vectors
         __m256 v0 = _mm256_loadu_ps( x );
@@ -1460,7 +1467,8 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
         // Compute the sum of the quants and set y[i].s
         const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
         const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
-        y[i].s = d * hsum_i32_8(_mm256_set_m128i(s1, s0));
+        y[i].s0 = d * hsum_i32_4(s0);
+        y[i].s1 = d * hsum_i32_4(s1);
 
         // Convert int32 to int16
         ni0 = _mm_packs_epi32( ni0, ni1 );
diff --git a/llama.cpp b/llama.cpp
index 00cce6e2a..4e92f5515 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -68,7 +68,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
         { MODEL_65B,   512ull * MB },
     };
     return _MEM_REQ_SCRATCH1;
-};
+}
 
 // 2*n_embd*n_ctx*n_layer*sizeof(float16)
 static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
@@ -80,7 +80,7 @@ static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
         { MODEL_65B,  5120ull * MB },
     };
     return _MEM_REQ_KV_SELF;
-};
+}
 
 // this is mostly needed for temporary mul_mat buffers to dequantize the data
 // not actually needed if BLAS is disabled
@@ -93,7 +93,7 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
         { MODEL_65B, 1536ull * MB },
     };
     return _MEM_REQ_EVAL;
-};
+}
 
 // default hparams (LLaMA 7B)
 struct llama_hparams {