ggml : portability fixes for VS 2017 (#12150)

* Add include files for std::min/max and std::toupper/tolower

* win32: move _USE_MATH_DEFINES before includes to ensure M_PI is defined

* Use GGML_RESTRICT instead of "restrict" keyword everywhere, and use "__restrict" in MSVC plain C mode

* win32: only use __restrict in MSVC if C11/C17 support is not enabled

---------

Co-authored-by: Marcus Groeber <Marcus.Groeber@cerence.com>

common/ngram-cache.cpp

@@ -7,6 +7,7 @@
 #include <cstdio>
 #include <fstream>
 #include <thread>
+#include <algorithm>
 
 void common_ngram_cache_update(common_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
                               std::vector<llama_token> & inp, int nnew, bool print_progress) {

common/sampling.cpp

@@ -4,6 +4,7 @@
 
 #include <cmath>
 #include <unordered_map>
+#include <algorithm>
 
 // the ring buffer works similarly to std::deque, but with a fixed capacity
 // TODO: deduplicate with llama-impl.h

common/speculative.cpp

@@ -5,6 +5,7 @@
 #include "sampling.h"
 
 #include <cstring>
+#include <algorithm>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5

examples/embedding/embedding.cpp

@@ -4,6 +4,7 @@
 #include "llama.h"
 
 #include <ctime>
+#include <algorithm>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data

examples/lookahead/lookahead.cpp

@@ -7,6 +7,7 @@
 #include <cstdio>
 #include <string>
 #include <vector>
+#include <algorithm>
 
 struct ngram_data {
     bool active = false;

examples/parallel/parallel.cpp

@@ -12,6 +12,7 @@
 #include <string>
 #include <vector>
 #include <ctime>
+#include <algorithm>
 
 // trim whitespace from the beginning and end of a string
 static std::string trim(const std::string & str) {

examples/passkey/passkey.cpp

@@ -7,6 +7,7 @@
 #include <cstdio>
 #include <string>
 #include <vector>
+#include <algorithm>
 
 static void print_usage(int, char ** argv) {
     LOG("\nexample usage:\n");

examples/quantize/quantize.cpp

@@ -8,6 +8,7 @@
 #include <unordered_map>
 #include <fstream>
 #include <cmath>
+#include <cctype>
 
 struct quant_option {
     std::string name;

examples/tts/tts.cpp

@@ -1,3 +1,5 @@
+#define _USE_MATH_DEFINES // For M_PI on MSVC
+
 #include "arg.h"
 #include "common.h"
 #include "sampling.h"
@@ -5,8 +7,6 @@
 #include "llama.h"
 #include "json.hpp"
 
-#define _USE_MATH_DEFINES // For M_PI on MSVC
-
 #include <algorithm>
 #include <cmath>
 #include <cstdio>

ggml/include/ggml.h

@@ -2140,7 +2140,11 @@ extern "C" {
 #        define GGML_RESTRICT
 #    endif
 #else
-#    define GGML_RESTRICT restrict
+#    if defined (_MSC_VER) && (__STDC_VERSION__ < 201112L)
+#        define GGML_RESTRICT __restrict
+#    else
+#        define GGML_RESTRICT restrict
+#    endif
 #endif
     typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);

ggml/src/ggml-backend-reg.cpp

@@ -8,6 +8,7 @@
 #include <string>
 #include <type_traits>
 #include <vector>
+#include <cctype>
 
 #ifdef _WIN32
 #    define WIN32_LEAN_AND_MEAN

ggml/src/ggml-backend.cpp

@@ -21,6 +21,7 @@
 #include <string.h>
 #include <string>
 #include <vector>
+#include <algorithm>
 
 #ifdef __APPLE__
 #include <sys/types.h>

ggml/src/ggml-cpu/ggml-cpu.c

@@ -247,9 +247,9 @@ typedef pthread_t ggml_thread_t;
 static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 
 
-static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc);
-static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc);
-static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc);
+static void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc);
+static void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc);
+static void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc);
 
 static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
     [GGML_TYPE_F32] = {
@@ -1451,7 +1451,7 @@ inline static void ggml_vec_div_f16 (const int n, ggml_fp16_t * z, const ggml_fp
     }
 }
 
-static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc) {
+static void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc) {
    assert(nrc == 1);
    UNUSED(nrc);
    UNUSED(bx);
@@ -1494,7 +1494,7 @@ static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float *
     *s = sumf;
 }
 
-static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc) {
+static void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc) {
     assert(nrc == 1);
     UNUSED(nrc);
     UNUSED(bx);
@@ -1562,7 +1562,7 @@ static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t
     *s = sumf;
 }
 
-static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc) {
+static void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc) {
     assert(nrc == 1);
     UNUSED(nrc);
     UNUSED(bx);
@@ -1606,10 +1606,10 @@ static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t *
 
 // compute GGML_VEC_DOT_UNROLL dot products at once
 // xs - x row stride in bytes
-inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * restrict s, void * restrict xv, ggml_fp16_t * restrict y) {
+inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GGML_RESTRICT s, void * GGML_RESTRICT xv, ggml_fp16_t * GGML_RESTRICT y) {
     ggml_float sumf[GGML_VEC_DOT_UNROLL] = { 0.0 };
 
-    ggml_fp16_t * restrict x[GGML_VEC_DOT_UNROLL];
+    ggml_fp16_t * GGML_RESTRICT x[GGML_VEC_DOT_UNROLL];
 
     for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
         x[i] = (ggml_fp16_t *) ((char *) xv + i*xs);
@@ -1659,7 +1659,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * re
     }
 }
 
-inline static void ggml_vec_mad_f32(const int n, float * restrict y, const float * restrict x, const float v) {
+inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const float * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD)
     const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -1690,7 +1690,7 @@ inline static void ggml_vec_mad_f32(const int n, float * restrict y, const float
 #endif
 }
 
-inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * restrict y, const ggml_fp16_t * restrict x, const float v) {
+inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD)
     const int np = (n & ~(GGML_F16_STEP - 1));
 
@@ -1722,10 +1722,10 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * restrict y, const
 }
 
 // xs and vs are byte strides of x and v
-inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * restrict y, const float * restrict xv, const float * restrict vv) {
+inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * GGML_RESTRICT y, const float * GGML_RESTRICT xv, const float * GGML_RESTRICT vv) {
 
-    const float * restrict x[GGML_VEC_MAD_UNROLL];
-    const float * restrict v[GGML_VEC_MAD_UNROLL];
+    const float * GGML_RESTRICT x[GGML_VEC_MAD_UNROLL];
+    const float * GGML_RESTRICT v[GGML_VEC_MAD_UNROLL];
 
     for (int i = 0; i < GGML_VEC_MAD_UNROLL; ++i) {
         x[i] = (const float *) ((const char *) xv + i*xs);

ggml/src/ggml-quants.c

@@ -28,7 +28,7 @@
 #define UNUSED GGML_UNUSED
 
 // reference implementation for deterministic creation of model files
-void quantize_row_q4_0_ref(const float * restrict x, block_q4_0 * restrict y, int64_t k) {
+void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_0;
 
     assert(k % qk == 0);
@@ -65,7 +65,7 @@ void quantize_row_q4_0_ref(const float * restrict x, block_q4_0 * restrict y, in
     }
 }
 
-void quantize_row_q4_1_ref(const float * restrict x, block_q4_1 * restrict y, int64_t k) {
+void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k) {
     const int qk = QK4_1;
 
     assert(k % qk == 0);
@@ -102,7 +102,7 @@ void quantize_row_q4_1_ref(const float * restrict x, block_q4_1 * restrict y, in
     }
 }
 
-void quantize_row_q5_0_ref(const float * restrict x, block_q5_0 * restrict y, int64_t k) {
+void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK5_0;
 
     assert(k % qk == 0);
@@ -146,7 +146,7 @@ void quantize_row_q5_0_ref(const float * restrict x, block_q5_0 * restrict y, in
     }
 }
 
-void quantize_row_q5_1_ref(const float * restrict x, block_q5_1 * restrict y, int64_t k) {
+void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k) {
     const int qk = QK5_1;
 
     assert(k % qk == 0);
@@ -191,7 +191,7 @@ void quantize_row_q5_1_ref(const float * restrict x, block_q5_1 * restrict y, in
 }
 
 // reference implementation for deterministic creation of model files
-void quantize_row_q8_0_ref(const float * restrict x, block_q8_0 * restrict y, int64_t k) {
+void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k) {
     assert(k % QK8_0 == 0);
     const int nb = k / QK8_0;
 
@@ -217,7 +217,7 @@ void quantize_row_q8_0_ref(const float * restrict x, block_q8_0 * restrict y, in
 }
 
 // reference implementation for deterministic creation of model files
-void quantize_row_q8_1_ref(const float * restrict x, block_q8_1 * restrict y, int64_t k) {
+void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k) {
     assert(QK8_1 == 32);
     assert(k % QK8_1 == 0);
     const int nb = k / QK8_1;
@@ -252,7 +252,7 @@ void quantize_row_q8_1_ref(const float * restrict x, block_q8_1 * restrict y, in
     }
 }
 
-void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_0;
 
     assert(k % qk == 0);
@@ -272,7 +272,7 @@ void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int6
     }
 }
 
-void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_1;
 
     assert(k % qk == 0);
@@ -293,7 +293,7 @@ void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int6
     }
 }
 
-void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK5_0;
 
     assert(k % qk == 0);
@@ -319,7 +319,7 @@ void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int6
     }
 }
 
-void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK5_1;
 
     assert(k % qk == 0);
@@ -346,7 +346,7 @@ void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int6
     }
 }
 
-void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK8_0;
 
     assert(k % qk == 0);
@@ -376,8 +376,8 @@ static inline int nearest_int(float fval) {
     return (i & 0x007fffff) - 0x00400000;
 }
 
-static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, int rmse_type,
-        const float * restrict qw) {
+static float make_qx_quants(int n, int nmax, const float * GGML_RESTRICT x, int8_t * GGML_RESTRICT L, int rmse_type,
+        const float * GGML_RESTRICT qw) {
     float max = 0;
     float amax = 0;
     for (int i = 0; i < n; ++i) {
@@ -445,7 +445,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
     return scale;
 }
 
-static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, bool do_rmse) {
+static float make_q3_quants(int n, int nmax, const float * GGML_RESTRICT x, int8_t * GGML_RESTRICT L, bool do_rmse) {
     float max = 0;
     float amax = 0;
     for (int i = 0; i < n; ++i) {
@@ -504,7 +504,7 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t *
     return 1/iscale;
 }
 
-static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min,
+static float make_qkx1_quants(int n, int nmax, const float * GGML_RESTRICT x, uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min,
         int ntry, float alpha) {
     float min = x[0];
     float max = x[0];
@@ -547,8 +547,8 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
     return scale;
 }
 
-static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
-        uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
+static float make_qkx2_quants(int n, int nmax, const float * GGML_RESTRICT x, const float * GGML_RESTRICT weights,
+        uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min, uint8_t * GGML_RESTRICT Laux,
         float rmin, float rdelta, int nstep, bool use_mad) {
     float min = x[0];
     float max = x[0];
@@ -628,7 +628,7 @@ static float make_qkx2_quants(int n, int nmax, const float * restrict x, const f
     return scale;
 }
 
-static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) {
+static inline void get_scale_min_k4(int j, const uint8_t * GGML_RESTRICT q, uint8_t * GGML_RESTRICT d, uint8_t * GGML_RESTRICT m) {
     if (j < 4) {
         *d = q[j] & 63; *m = q[j + 4] & 63;
     } else {
@@ -639,7 +639,7 @@ static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t *
 
 //========================- 2-bit (de)-quantization
 
-void quantize_row_q2_K_ref(const float * restrict x, block_q2_K * restrict y, int64_t k) {
+void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
@@ -709,7 +709,7 @@ void quantize_row_q2_K_ref(const float * restrict x, block_q2_K * restrict y, in
     }
 }
 
-void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
@@ -741,8 +741,8 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int6
     }
 }
 
-static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
-        uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
+static float make_qkx3_quants(int n, int nmax, const float * GGML_RESTRICT x, const float * GGML_RESTRICT weights,
+        uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min, uint8_t * GGML_RESTRICT Laux,
         float rmin, float rdelta, int nstep, bool use_mad) {
     float min = x[0];
     float max = x[0];
@@ -824,7 +824,7 @@ static float make_qkx3_quants(int n, int nmax, const float * restrict x, const f
     return scale;
 }
 
-static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, const float * quant_weights) {
+static float make_qp_quants(int n, int nmax, const float * GGML_RESTRICT x, uint8_t * GGML_RESTRICT L, const float * quant_weights) {
     float max = 0;
     for (int i = 0; i < n; ++i) {
         max = MAX(max, x[i]);
@@ -897,7 +897,7 @@ static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t *
     return sumlx/suml2;
 }
 
-static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restrict y, int k, const float * restrict quant_weights) {
+static void quantize_row_q2_K_impl(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int k, const float * GGML_RESTRICT quant_weights) {
     GGML_ASSERT(quant_weights);
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
@@ -917,7 +917,7 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri
         for (int j = 0; j < QK_K; ++j) sumx2 += x[j]*x[j];
         float sigma2 = sumx2/QK_K;
         for (int j = 0; j < QK_K/16; ++j) {
-            const float * restrict qw = quant_weights + QK_K * i + 16*j;
+            const float * GGML_RESTRICT qw = quant_weights + QK_K * i + 16*j;
             for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j + l]*x[16*j + l]);
             for (int l = 0; l < QK_K/16; ++l) sw[j] += weight[l];
             scales[j] = make_qkx3_quants(16, 3, x + 16*j, weight, L + 16*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
@@ -959,7 +959,7 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri
     }
 }
 
-size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     size_t row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row);
     if (!quant_weights) {
         quantize_row_q2_K_ref(src, dst, (int64_t)nrow*n_per_row);
@@ -977,7 +977,7 @@ size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nr
 
 //========================= 3-bit (de)-quantization
 
-void quantize_row_q3_K_ref(const float * restrict x, block_q3_K * restrict y, int64_t k) {
+void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
@@ -1053,7 +1053,7 @@ void quantize_row_q3_K_ref(const float * restrict x, block_q3_K * restrict y, in
     }
 }
 
-void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
@@ -1067,8 +1067,8 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int6
 
         const float d_all = GGML_FP16_TO_FP32(x[i].d);
 
-        const uint8_t * restrict q = x[i].qs;
-        const uint8_t * restrict hm = x[i].hmask;
+        const uint8_t * GGML_RESTRICT q = x[i].qs;
+        const uint8_t * GGML_RESTRICT hm = x[i].hmask;
         uint8_t m = 1;
 
         memcpy(aux, x[i].scales, 12);
@@ -1103,7 +1103,7 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int6
     }
 }
 
-static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int64_t n_per_row, const float * restrict quant_weights) {
+static void quantize_row_q3_K_impl(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t n_per_row, const float * GGML_RESTRICT quant_weights) {
     assert(n_per_row % QK_K == 0);
     const int nb = n_per_row / QK_K;
 
@@ -1187,7 +1187,7 @@ static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restri
     }
 }
 
-size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     size_t row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row);
     if (!quant_weights) {
         quantize_row_q3_K_ref(src, dst, (int64_t)nrow*n_per_row);
@@ -1205,7 +1205,7 @@ size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nr
 
 // ====================== 4-bit (de)-quantization
 
-void quantize_row_q4_K_ref(const float * restrict x, block_q4_K * restrict y, int64_t k) {
+void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
@@ -1277,7 +1277,7 @@ void quantize_row_q4_K_ref(const float * restrict x, block_q4_K * restrict y, in
     }
 }
 
-void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
@@ -1301,7 +1301,7 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int6
     }
 }
 
-static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int64_t n_per_row, const float * quant_weights) {
+static void quantize_row_q4_K_impl(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
     assert(n_per_row % QK_K == 0);
     const int64_t nb = n_per_row / QK_K;
 
@@ -1374,7 +1374,7 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri
     }
 }
 
-size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     size_t row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row);
     if (!quant_weights) {
         quantize_row_q4_K_ref(src, dst, (int64_t)nrow*n_per_row);
@@ -1392,7 +1392,7 @@ size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nr
 
 // ====================== 5-bit (de)-quantization
 
-void quantize_row_q5_K_ref(const float * restrict x, block_q5_K * restrict y, int64_t k) {
+void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -1454,8 +1454,8 @@ void quantize_row_q5_K_ref(const float * restrict x, block_q5_K * restrict y, in
             }
         }
 
-        uint8_t * restrict qh = y[i].qh;
-        uint8_t * restrict ql = y[i].qs;
+        uint8_t * GGML_RESTRICT qh = y[i].qh;
+        uint8_t * GGML_RESTRICT ql = y[i].qs;
         memset(qh, 0, QK_K/8);
 
         uint8_t m1 = 1, m2 = 2;
@@ -1479,7 +1479,7 @@ void quantize_row_q5_K_ref(const float * restrict x, block_q5_K * restrict y, in
     }
 }
 
-void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -1506,7 +1506,7 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int6
     }
 }
 
-static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int64_t n_per_row, const float * quant_weights) {
+static void quantize_row_q5_K_impl(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
     assert(n_per_row % QK_K == 0);
     const int64_t nb = n_per_row / QK_K;
 
@@ -1573,8 +1573,8 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri
             }
         }
 
-        uint8_t * restrict qh = y[i].qh;
-        uint8_t * restrict ql = y[i].qs;
+        uint8_t * GGML_RESTRICT qh = y[i].qh;
+        uint8_t * GGML_RESTRICT ql = y[i].qs;
         memset(qh, 0, QK_K/8);
 
         uint8_t m1 = 1, m2 = 2;
@@ -1599,7 +1599,7 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri
     }
 }
 
-size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     size_t row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row);
     if (!quant_weights) {
         quantize_row_q5_K_ref(src, dst, (int64_t)nrow*n_per_row);
@@ -1617,7 +1617,7 @@ size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nr
 
 // ====================== 6-bit (de)-quantization
 
-void quantize_row_q6_K_ref(const float * restrict x, block_q6_K * restrict y, int64_t k) {
+void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -1667,8 +1667,8 @@ void quantize_row_q6_K_ref(const float * restrict x, block_q6_K * restrict y, in
             }
         }
 
-        uint8_t * restrict ql = y[i].ql;
-        uint8_t * restrict qh = y[i].qh;
+        uint8_t * GGML_RESTRICT ql = y[i].ql;
+        uint8_t * GGML_RESTRICT qh = y[i].qh;
         for (int j = 0; j < QK_K; j += 128) {
             for (int l = 0; l < 32; ++l) {
                 const uint8_t q1 = L[j + l +  0] & 0xF;
@@ -1687,16 +1687,16 @@ void quantize_row_q6_K_ref(const float * restrict x, block_q6_K * restrict y, in
     }
 }
 
-void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
     for (int i = 0; i < nb; i++) {
         const float d = GGML_FP16_TO_FP32(x[i].d);
 
-        const uint8_t * restrict ql = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict sc = x[i].scales;
+        const uint8_t * GGML_RESTRICT ql = x[i].ql;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const int8_t  * GGML_RESTRICT sc = x[i].scales;
 
         for (int n = 0; n < QK_K; n += 128) {
             for (int l = 0; l < 32; ++l) {
@@ -1718,7 +1718,7 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int6
     }
 }
 
-static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int64_t n_per_row, const float * quant_weights) {
+static void quantize_row_q6_K_impl(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
     assert(n_per_row % QK_K == 0);
     const int64_t nb = n_per_row / QK_K;
 
@@ -1781,8 +1781,8 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri
             }
         }
 
-        uint8_t * restrict ql = y[i].ql;
-        uint8_t * restrict qh = y[i].qh;
+        uint8_t * GGML_RESTRICT ql = y[i].ql;
+        uint8_t * GGML_RESTRICT qh = y[i].qh;
         for (int j = 0; j < QK_K; j += 128) {
             for (int l = 0; l < 32; ++l) {
                 const uint8_t q1 = L[j + l +  0] & 0xF;
@@ -1802,7 +1802,7 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri
     }
 }
 
-size_t quantize_q6_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     size_t row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row);
     if (!quant_weights) {
         quantize_row_q6_K_ref(src, dst, (int64_t)nrow*n_per_row);
@@ -1818,7 +1818,7 @@ size_t quantize_q6_K(const float * restrict src, void * restrict dst, int64_t nr
     return nrow * row_size;
 }
 
-static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restrict y, int64_t n_per_row, const float * quant_weights) {
+static void quantize_row_q4_0_impl(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
     static_assert(QK4_0 == 32, "QK4_0 must be 32");
 
     if (!quant_weights) {
@@ -1846,7 +1846,7 @@ static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restri
     }
 }
 
-size_t quantize_q4_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     if (!quant_weights) {
         quantize_row_q4_0_ref(src, dst, (int64_t)nrow*n_per_row);
         return nrow * ggml_row_size(GGML_TYPE_Q4_0, n_per_row);
@@ -1861,7 +1861,7 @@ size_t quantize_q4_0(const float * restrict src, void * restrict dst, int64_t nr
     return nrow * row_size;
 }
 
-static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restrict y, int64_t n_per_row, const float * quant_weights) {
+static void quantize_row_q4_1_impl(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
     static_assert(QK4_1 == 32, "QK4_1 must be 32");
 
     if (!quant_weights) {
@@ -1891,7 +1891,7 @@ static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restri
     }
 }
 
-size_t quantize_q4_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     if (!quant_weights) {
         quantize_row_q4_1_ref(src, dst, (int64_t)nrow*n_per_row);
         return nrow * ggml_row_size(GGML_TYPE_Q4_1, n_per_row);
@@ -1906,7 +1906,7 @@ size_t quantize_q4_1(const float * restrict src, void * restrict dst, int64_t nr
     return nrow * row_size;
 }
 
-static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restrict y, int64_t n_per_row, const float * quant_weights) {
+static void quantize_row_q5_0_impl(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
     static_assert(QK5_0 == 32, "QK5_0 must be 32");
 
     if (!quant_weights) {
@@ -1945,7 +1945,7 @@ static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restri
     }
 }
 
-size_t quantize_q5_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     if (!quant_weights) {
         quantize_row_q5_0_ref(src, dst, (int64_t)nrow*n_per_row);
         return nrow * ggml_row_size(GGML_TYPE_Q5_0, n_per_row);
@@ -1960,7 +1960,7 @@ size_t quantize_q5_0(const float * restrict src, void * restrict dst, int64_t nr
     return nrow * row_size;
 }
 
-static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restrict y, int64_t n_per_row, const float * quant_weights) {
+static void quantize_row_q5_1_impl(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
     static_assert(QK5_1 == 32, "QK5_1 must be 32");
 
     if (!quant_weights) {
@@ -1998,7 +1998,7 @@ static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restri
     }
 }
 
-size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     if (!quant_weights) {
         quantize_row_q5_1_ref(src, dst, (int64_t)nrow*n_per_row);
         return nrow * ggml_row_size(GGML_TYPE_Q5_1, n_per_row);
@@ -2013,7 +2013,7 @@ size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nr
     return nrow * row_size;
 }
 
-size_t quantize_q8_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     (void)quant_weights; // not used
     const size_t row_size = ggml_row_size(GGML_TYPE_Q8_0, n_per_row);
     quantize_row_q8_0_ref(src, dst, (int64_t)nrow*n_per_row);
@@ -2022,7 +2022,7 @@ size_t quantize_q8_0(const float * restrict src, void * restrict dst, int64_t nr
 
 // ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
 
-void quantize_row_tq1_0_ref(const float * restrict x, block_tq1_0 * restrict y, int64_t k) {
+void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2088,7 +2088,7 @@ void quantize_row_tq1_0_ref(const float * restrict x, block_tq1_0 * restrict y,
     }
 }
 
-void quantize_row_tq2_0_ref(const float * restrict x, block_tq2_0 * restrict y, int64_t k) {
+void quantize_row_tq2_0_ref(const float * GGML_RESTRICT x, block_tq2_0 * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2120,21 +2120,21 @@ void quantize_row_tq2_0_ref(const float * restrict x, block_tq2_0 * restrict y,
     }
 }
 
-size_t quantize_tq1_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_tq1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     (void)quant_weights; // not used
     const size_t row_size = ggml_row_size(GGML_TYPE_TQ1_0, n_per_row);
     quantize_row_tq1_0_ref(src, dst, (int64_t)nrow*n_per_row);
     return nrow * row_size;
 }
 
-size_t quantize_tq2_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_tq2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     (void)quant_weights; // not used
     const size_t row_size = ggml_row_size(GGML_TYPE_TQ2_0, n_per_row);
     quantize_row_tq2_0_ref(src, dst, (int64_t)nrow*n_per_row);
     return nrow * row_size;
 }
 
-void dequantize_row_tq1_0(const block_tq1_0 * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_tq1_0(const block_tq1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2173,7 +2173,7 @@ void dequantize_row_tq1_0(const block_tq1_0 * restrict x, float * restrict y, in
     }
 }
 
-void dequantize_row_tq2_0(const block_tq2_0 * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_tq2_0(const block_tq2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2194,7 +2194,7 @@ void dequantize_row_tq2_0(const block_tq2_0 * restrict x, float * restrict y, in
 
 // ====================== "True" 2-bit (de)-quantization
 
-void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2222,7 +2222,7 @@ void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y
 
 // ====================== 2.3125 bpw (de)-quantization
 
-void dequantize_row_iq2_xs(const block_iq2_xs * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_iq2_xs(const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2249,7 +2249,7 @@ void dequantize_row_iq2_xs(const block_iq2_xs * restrict x, float * restrict y,
 
 // ====================== 2.5625 bpw (de)-quantization
 
-void dequantize_row_iq2_s(const block_iq2_s * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_iq2_s(const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2281,7 +2281,7 @@ void dequantize_row_iq2_s(const block_iq2_s * restrict x, float * restrict y, in
 
 // ====================== 3.0625 bpw (de)-quantization
 
-void dequantize_row_iq3_xxs(const block_iq3_xxs * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2313,7 +2313,7 @@ void dequantize_row_iq3_xxs(const block_iq3_xxs * restrict x, float * restrict y
 
 // ====================== 3.3125 bpw (de)-quantization
 
-void dequantize_row_iq3_s(const block_iq3_s * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_iq3_s(const block_iq3_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2356,7 +2356,7 @@ void dequantize_row_iq3_s(const block_iq3_s * restrict x, float * restrict y, in
 
 // ====================== 1.5625 bpw (de)-quantization
 
-void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_iq1_s(const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2381,7 +2381,7 @@ void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, in
     }
 }
 
-void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_iq1_m(const block_iq1_m * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2433,7 +2433,7 @@ void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, in
 
 static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
 
-void dequantize_row_iq4_nl(const block_iq4_nl * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_iq4_nl(const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK4_NL == 0);
     const int64_t nb = k / QK4_NL;
 
@@ -2451,7 +2451,7 @@ void dequantize_row_iq4_nl(const block_iq4_nl * restrict x, float * restrict y,
     }
 }
 
-void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_iq4_xs(const block_iq4_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2476,7 +2476,7 @@ void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y,
 
 //===================================== Q8_K ==============================================
 
-void quantize_row_q8_K_ref(const float * restrict x, block_q8_K * restrict y, int64_t k) {
+void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2515,7 +2515,7 @@ void quantize_row_q8_K_ref(const float * restrict x, block_q8_K * restrict y, in
     }
 }
 
-void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int64_t k) {
+void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
 
@@ -2927,8 +2927,8 @@ void iq2xs_free_impl(enum ggml_type type) {
     }
 }
 
-static int iq2_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid,
-        const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) {
+static int iq2_find_best_neighbour(const uint16_t * GGML_RESTRICT neighbours, const uint64_t * GGML_RESTRICT grid,
+        const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float scale, int8_t * GGML_RESTRICT L) {
     int num_neighbors = neighbours[0];
     GGML_ASSERT(num_neighbors > 0);
     float best_d2 = FLT_MAX;
@@ -2951,7 +2951,7 @@ static int iq2_find_best_neighbour(const uint16_t * restrict neighbours, const u
     return grid_index;
 }
 
-static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) {
+static void quantize_row_iq2_xxs_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights) {
 
     const int gindex = iq2_data_index(GGML_TYPE_IQ2_XXS);
 
@@ -3124,7 +3124,7 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
     }
 }
 
-static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) {
+static void quantize_row_iq2_xs_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights) {
 
     const int gindex = iq2_data_index(GGML_TYPE_IQ2_XS);
 
@@ -3304,7 +3304,7 @@ static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict v
     }
 }
 
-size_t quantize_iq2_xxs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK_K == 0);
     int64_t nblock = n_per_row/QK_K;
     char * qrow = (char *)dst;
@@ -3316,7 +3316,7 @@ size_t quantize_iq2_xxs(const float * restrict src, void * restrict dst, int64_t
     return nrow * nblock * sizeof(block_iq2_xxs);
 }
 
-size_t quantize_iq2_xs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_iq2_xs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK_K == 0);
     int64_t nblock = n_per_row/QK_K;
     char * qrow = (char *)dst;
@@ -3521,8 +3521,8 @@ void iq3xs_free_impl(int grid_size) {
     }
 }
 
-static int iq3_find_best_neighbour(const uint16_t * restrict neighbours, const uint32_t * restrict grid,
-        const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) {
+static int iq3_find_best_neighbour(const uint16_t * GGML_RESTRICT neighbours, const uint32_t * GGML_RESTRICT grid,
+        const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float scale, int8_t * GGML_RESTRICT L) {
     int num_neighbors = neighbours[0];
     GGML_ASSERT(num_neighbors > 0);
     float best_d2 = FLT_MAX;
@@ -3545,8 +3545,8 @@ static int iq3_find_best_neighbour(const uint16_t * restrict neighbours, const u
     return grid_index;
 }
 
-static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, void * restrict vy, int64_t n,
-        const float * restrict quant_weights) {
+static void quantize_row_iq3_xxs_impl(int grid_size, const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n,
+        const float * GGML_RESTRICT quant_weights) {
 
     const int gindex = iq3_data_index(grid_size);
 
@@ -3758,7 +3758,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, v
     }
 }
 
-size_t quantize_iq3_xxs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK_K == 0);
     int64_t nblock = n_per_row/QK_K;
     char * qrow = (char *)dst;
@@ -3770,13 +3770,13 @@ size_t quantize_iq3_xxs(const float * restrict src, void * restrict dst, int64_t
     return nrow * nblock * sizeof(block_iq3_xxs);
 }
 
-void quantize_row_iq3_xxs_ref(const float * restrict x, block_iq3_xxs * restrict y, int64_t k) {
+void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_row_iq3_xxs_impl(256, x, y, k, NULL);
 }
 
-static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, void * restrict vy, int n,
-        const float * restrict quant_weights,
+static void quantize_row_iq3_s_impl(int block_size, const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int n,
+        const float * GGML_RESTRICT quant_weights,
         float   * scales,
         float   * weight,
         float   * xval,
@@ -3958,7 +3958,7 @@ static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, vo
 }
 
 #define IQ3S_BLOCK_SIZE 32
-size_t quantize_iq3_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_iq3_s(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK_K == 0);
     int64_t nblock = n_per_row/QK_K;
     float scales[QK_K/IQ3S_BLOCK_SIZE];
@@ -3980,7 +3980,7 @@ size_t quantize_iq3_s(const float * restrict src, void * restrict dst, int64_t n
     return nrow * nblock * sizeof(block_iq3_s);
 }
 
-void quantize_row_iq3_s_ref(const float * restrict x, block_iq3_s * restrict y, int64_t k) {
+void quantize_row_iq3_s_ref(const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_iq3_s(x, y, 1, k, NULL);
 }
@@ -3988,8 +3988,8 @@ void quantize_row_iq3_s_ref(const float * restrict x, block_iq3_s * restrict y,
 
 // =================================== 1.5 bpw ===================================================
 
-static int iq1_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid,
-        const float * restrict xval, const float * restrict weight, float * scale, int8_t * restrict L, int ngrid) {
+static int iq1_find_best_neighbour(const uint16_t * GGML_RESTRICT neighbours, const uint64_t * GGML_RESTRICT grid,
+        const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float * scale, int8_t * GGML_RESTRICT L, int ngrid) {
     int num_neighbors = neighbours[0];
     GGML_ASSERT(num_neighbors > 0);
     float best_score = -FLT_MAX;
@@ -4048,8 +4048,8 @@ static int iq1_find_best_neighbour(const uint16_t * restrict neighbours, const u
     return grid_index;
 }
 
-static int iq1_find_best_neighbour2(const uint16_t * restrict neighbours, const uint64_t * restrict grid,
-        const float * restrict xval, const float * restrict weight, float scale, const float * restrict xg, int8_t * restrict L, int ngrid) {
+static int iq1_find_best_neighbour2(const uint16_t * GGML_RESTRICT neighbours, const uint64_t * GGML_RESTRICT grid,
+        const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float scale, const float * GGML_RESTRICT xg, int8_t * GGML_RESTRICT L, int ngrid) {
     int num_neighbors = neighbours[0];
     GGML_ASSERT(num_neighbors > 0);
     float best_score = FLT_MAX;
@@ -4113,7 +4113,7 @@ static int iq1_sort_helper(const void * left, const void * right) {
 
 #define IQ1S_BLOCK_SIZE 32
 #define IQ1M_BLOCK_SIZE 16
-static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights,
+static void quantize_row_iq1_s_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights,
         float    * scales,
         float    * weight,
         float    * sumx,
@@ -4271,7 +4271,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
     }
 }
 
-size_t quantize_iq1_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_iq1_s(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK_K == 0);
     float  scales[QK_K/IQ1S_BLOCK_SIZE];
     float  weight[IQ1S_BLOCK_SIZE];
@@ -4291,7 +4291,7 @@ size_t quantize_iq1_s(const float * restrict src, void * restrict dst, int64_t n
     return nrow * nblock * sizeof(block_iq1_s);
 }
 
-static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights,
+static void quantize_row_iq1_m_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights,
         float    * scales,
         float    * weight,
         float    * pairs,
@@ -4539,7 +4539,7 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
     }
 }
 
-size_t quantize_iq1_m(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_iq1_m(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK_K == 0);
     float  scales[QK_K/IQ1M_BLOCK_SIZE];
     float  weight[IQ1M_BLOCK_SIZE];
@@ -4570,7 +4570,7 @@ static inline int best_index_int8(int n, const int8_t * val, float x) {
     return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
 }
 
-static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * restrict x,
+static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * GGML_RESTRICT x,
         ggml_fp16_t * dh, uint8_t * q4, uint16_t * scales_h, uint8_t * scales_l,
         float * scales, float * weight, uint8_t * L,
         const int8_t * values,
@@ -4681,7 +4681,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
     }
 }
 
-size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_iq4_nl(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK4_NL == 0);
     int64_t nblock = n_per_row/QK4_NL;
     char * qrow = (char *)dst;
@@ -4703,8 +4703,8 @@ size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int64_t
     return nrow * nblock * sizeof(block_iq4_nl);
 }
 
-//void quantize_row_iq4_nl_ref(const float * restrict x, void * restrict vy, int64_t k) {
-void quantize_row_iq4_nl_ref(const float * restrict x, block_iq4_nl * restrict y, int64_t k) {
+//void quantize_row_iq4_nl_ref(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+void quantize_row_iq4_nl_ref(const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k) {
     GGML_ASSERT(k%QK4_NL == 0);
     int64_t nblock = k/QK4_NL;
     uint8_t L[QK4_NL];
@@ -4719,7 +4719,7 @@ void quantize_row_iq4_nl_ref(const float * restrict x, block_iq4_nl * restrict y
     }
 }
 
-size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_iq4_xs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK_K == 0);
     int64_t nblock = n_per_row/QK_K;
     char * qrow = (char *)dst;
@@ -4739,14 +4739,14 @@ size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t
     return nrow * nblock * sizeof(block_iq4_xs);
 }
 
-void quantize_row_iq4_xs_ref(const float * restrict x, block_iq4_xs * restrict y, int64_t k) {
+void quantize_row_iq4_xs_ref(const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_iq4_xs(x, y, 1, k, NULL);
 }
 
 // =============================== 2.5625 bpw
 
-static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) {
+static void quantize_row_iq2_s_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights) {
 
     const int gindex = iq2_data_index(GGML_TYPE_IQ2_S);
 
@@ -4914,7 +4914,7 @@ static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy
     }
 }
 
-size_t quantize_iq2_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+size_t quantize_iq2_s(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK_K == 0);
     int64_t nblock = n_per_row/QK_K;
     char * qrow = (char *)dst;
@@ -4926,7 +4926,7 @@ size_t quantize_iq2_s(const float * restrict src, void * restrict dst, int64_t n
     return nrow * nblock * sizeof(block_iq2_s);
 }
 
-void quantize_row_iq2_s_ref(const float * restrict x, block_iq2_s * restrict y, int64_t k) {
+void quantize_row_iq2_s_ref(const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_iq2_s(x, y, 1, k, NULL);
 }

ggml/src/ggml.c

@@ -565,9 +565,9 @@ FILE * ggml_fopen(const char * fname, const char * mode) {
 #endif
 
 }
-static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc);
-static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc);
-static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc);
+static void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc);
+static void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc);
+static void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc);
 
 static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
     [GGML_TYPE_I8] = {

src/llama-chat.cpp

@@ -4,6 +4,7 @@
 
 #include <map>
 #include <sstream>
+#include <algorithm>
 
 #if __cplusplus >= 202000L
     #define LU8(x) (const char*)(u8##x)

src/llama-kv-cache.h

@@ -6,6 +6,7 @@
 
 #include <set>
 #include <vector>
+#include <algorithm>
 
 struct llama_kv_cell {
     llama_pos pos   = -1;

src/llama-mmap.cpp

@@ -8,6 +8,7 @@
 #include <climits>
 #include <stdexcept>
 #include <cerrno>
+#include <algorithm>
 
 #ifdef __has_include
     #if __has_include(<unistd.h>)

src/llama-vocab.cpp

@@ -16,6 +16,7 @@
 #include <queue>
 #include <set>
 #include <unordered_map>
+#include <cctype>
 
 //
 // helpers