From 785829dfe8baf0213f2ff66963d28c62f92d7930 Mon Sep 17 00:00:00 2001
From: Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Date: Thu, 20 Jul 2023 15:18:43 +0300
Subject: [PATCH 01/20] Faster Q4_K on Metal (#2290)

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
---
 ggml-metal.m     |   7 +-
 ggml-metal.metal | 262 ++++++++++++++++++++++++++++-------------------
 2 files changed, 160 insertions(+), 109 deletions(-)

diff --git a/ggml-metal.m b/ggml-metal.m
index d80a380d7..5e2a21100 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -694,8 +694,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
                                         } break;
                                     case GGML_TYPE_Q5_K:
@@ -739,7 +739,8 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
 
-                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
+                                    src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q2_K ||
diff --git a/ggml-metal.metal b/ggml-metal.metal
index ee56336ac..a9d134d6e 100644
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
@@ -1452,6 +1452,7 @@ kernel void kernel_mul_mat_q3_K_f32(
 
 }
 
+#if QK_K == 256
 kernel void kernel_mul_mat_q4_K_f32(
         device const  void * src0,
         device const float * src1,
@@ -1459,131 +1460,180 @@ kernel void kernel_mul_mat_q4_K_f32(
         constant   int64_t & ne00,
         constant   int64_t & ne10,
         constant   int64_t & ne0,
-        threadgroup float  * sum [[threadgroup(0)]],
+        constant   int64_t & ne01[[buffer(4)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2 tpitg[[thread_position_in_threadgroup]],
-        uint2  tptg[[threads_per_threadgroup]]) {
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-
-    const int nth = tptg.x*tptg.y;
-    const int ith = tptg.y*tpitg.x + tpitg.y;
-
-    device const block_q4_K * x = (device const block_q4_K *) src0 + r0*nb;
-    device const float     * yy = (device const float      *) src1 + r1*ne10;
-
-    float sumf = 0;
-
-#if QK_K == 256
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
 
-    const int tid = tpitg.y;   // 0...16
-    const int il  = tid/4;     // 0...3
-    const int ir  = tid - 4*il;// 0...3
-    const int n   = 4;
+    const int ix = tiisg/8;  // 0...3
+    const int it = tiisg%8;  // 0...7
+    const int im = it/4;     // 0 or 1
+    const int ir = it%4;     // 0...3
 
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+    device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row;
+    device const float      * y = (device const float      *) src1 + r1*ne10;
+    float yl[16];
+    float yh[16];
+    float sumf[N_DST]={0.f}, all_sum;
 
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
+    const int step = sizeof(block_q4_K) * nb / 2;
 
-    uchar2 sc1, sc2, sc3, sc4;
+    device const float * y4 = y + ix * QK_K + 64 * im + 8 * ir;
 
-    for (int i = tpitg.x; i < nb; i += tptg.x) {
+    uint16_t sc16[4];
+    thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
 
-        device const uint8_t * q1 = (x + i)->qs + q_offset;
-        device const uint8_t * q2 = q1 + 64;
-        device const float   * y1 = yy + i*QK_K + y_offset;
-        device const float   * y2 = y1 + 128;
-
-        const float dall = (float)((x + i)->d);
-        const float dmin = (float)((x + i)->dmin);
-
-        device const uint16_t * a = (device const uint16_t *)(x + i)->scales;
-        sc1 = as_type<uchar2>((uint16_t)(a[im+0] & kmask1));
-        sc2 = as_type<uchar2>((uint16_t)(a[im+2] & kmask1));
-        sc3 = as_type<uchar2>((uint16_t)(((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2)));
-        sc4 = as_type<uchar2>((uint16_t)(((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2)));
-
-        float4 s = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < n; ++l) {
-
-            s[0] += y1[l] * (q1[l] & 0xF); s[1] += y1[l+32] * (q1[l] >> 4);
-            s[2] += y2[l] * (q2[l] & 0xF); s[3] += y2[l+32] * (q2[l] >> 4);
-            smin += y1[l] * sc2[0] + y1[l+32] * sc2[1] + y2[l] * sc4[0] + y2[l+32] * sc4[1];
+    for (int ib = ix; ib < nb; ib += 4) {
 
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+0] = y4[i+  0]; sumy[0] += yl[i+0];
+            yl[i+8] = y4[i+ 32]; sumy[1] += yl[i+8];
+            yh[i+0] = y4[i+128]; sumy[2] += yh[i+0];
+            yh[i+8] = y4[i+160]; sumy[3] += yh[i+8];
         }
-        sumf += dall * (s[0] * sc1[0] + s[1] * sc1[1] + s[2] * sc3[0] + s[3] * sc3[1]) - dmin * smin;
 
+        device const uint16_t * sc = (device const uint16_t *)x[ib].scales + im;
+        device const uint16_t * q1 = (device const uint16_t *)x[ib].qs + 16 * im + 4 * ir;
+        device const half     * dh = &x[ib].d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            sc16[0] = sc[0] & kmask1;
+            sc16[1] = sc[2] & kmask1;
+            sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
+            sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);
+
+            device const uint16_t * q2 = q1 + 32;
+
+            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
+            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                acc1[0] += yl[i+0] * (q1[i/2] & 0x000F);
+                acc1[1] += yl[i+1] * (q1[i/2] & 0x0F00);
+                acc1[2] += yl[i+8] * (q1[i/2] & 0x00F0);
+                acc1[3] += yl[i+9] * (q1[i/2] & 0xF000);
+                acc2[0] += yh[i+0] * (q2[i/2] & 0x000F);
+                acc2[1] += yh[i+1] * (q2[i/2] & 0x0F00);
+                acc2[2] += yh[i+8] * (q2[i/2] & 0x00F0);
+                acc2[3] += yh[i+9] * (q2[i/2] & 0xF000);
+            }
+
+            float dall = dh[0];
+            float dmin = dh[1];
+            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8[0] +
+                                 (acc1[2] + 1.f/256.f * acc1[3]) * sc8[1] * 1.f/16.f +
+                                 (acc2[0] + 1.f/256.f * acc2[1]) * sc8[4] +
+                                 (acc2[2] + 1.f/256.f * acc2[3]) * sc8[5] * 1.f/16.f) -
+                         dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
+
+            q1 += step;
+            sc += step;
+            dh += step;
+        }
+
+        y4 += 4 * QK_K;
     }
-#else
-    uint16_t aux16[2];
-    thread const uint8_t * scales = (thread const uint8_t *)aux16;
 
-    const int il  = 4*tpitg.x;
-
-    for (int i = tpitg.y; i < nb; i += tptg.y) {
-
-        device const uint8_t * q = x[i].qs + il;
-        device const float   * y = yy + i * QK_K + il;
-
-        const float d = (float)x[i].d[0];
-        const float m = (float)x[i].d[1];
-
-        device const uint16_t * a = (device const uint16_t *)x[i].scales;
-        aux16[0] = a[0] & 0x0f0f;
-        aux16[1] = (a[0] >> 4) & 0x0f0f;
-
-        for (int l = 0; l < 4; ++l) {
-            sumf += d * scales[0] * (y[l+ 0] * (q[l] & 0xF) + y[l+16] * (q[l+16] & 0xF)) - m * scales[2] * (y[l+ 0] + y[l+16])
-                  + d * scales[1] * (y[l+32] * (q[l] >>  4) + y[l+48] * (q[l+16] >>  4)) - m * scales[3] * (y[l+32] + y[l+48]);
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + first_row + row] = all_sum;
         }
     }
-#endif
-
-    sum[ith] = sumf;
-
-    //
-    // Accumulate the sum from all threads in the threadgroup
-    // This version is slightly faster than the commented out one below,
-    // which I copy-pasted from ggerganov's q4_0 dot product for metal.
-    //
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%4 == 0) {
-        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%16 == 0) {
-        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith == 0) {
-        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
-        dst[r1*ne0 + r0] = sum[0];
-    }
-
-    //// accumulate the sum from all threads in the threadgroup
-    //threadgroup_barrier(mem_flags::mem_threadgroup);
-    //for (uint i = nth/2; i > 0; i /= 2) {
-    //    if (ith < i) {
-    //        sum[ith] += sum[ith + i];
-    //    }
-    //    threadgroup_barrier(mem_flags::mem_threadgroup);
-    //}
-
-    //if (ith == 0) {
-    //    dst[r1*ne0 + r0] = sum[0];
-    //}
 }
+#else
+kernel void kernel_mul_mat_q4_K_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne10,
+        constant   int64_t & ne0,
+        constant   int64_t & ne01[[buffer(4)]],
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int ix = tiisg/4;  // 0...7
+    const int it = tiisg%4;  // 0...3
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+    device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row;
+    device const float      * y = (device const float      *) src1 + r1*ne10;
+    float yl[8];
+    float yh[8];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int step = sizeof(block_q4_K) * nb / 2;
+
+    device const float * y4 = y + ix * QK_K + 8 * it;
+
+    uint16_t sc16[4];
+
+    for (int ib = ix; ib < nb; ib += 8) {
+
+        float2 sumy = {0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i] = y4[i+ 0]; sumy[0] += yl[i];
+            yh[i] = y4[i+32]; sumy[1] += yh[i];
+        }
+
+        device const uint16_t * sc = (device const uint16_t *)x[ib].scales;
+        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
+        device const half     * dh = x[ib].d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            sc16[0] = sc[0] & 0x000f;
+            sc16[1] = sc[0] & 0x0f00;
+            sc16[2] = sc[0] & 0x00f0;
+            sc16[3] = sc[0] & 0xf000;
+
+            float2 acc1 = {0.f, 0.f};
+            float2 acc2 = {0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                acc1[0] += yl[i+0] * (qs[i/2] & 0x000F);
+                acc1[1] += yl[i+1] * (qs[i/2] & 0x0F00);
+                acc2[0] += yh[i+0] * (qs[i/2] & 0x00F0);
+                acc2[1] += yh[i+1] * (qs[i/2] & 0xF000);
+            }
+
+            float dall = dh[0];
+            float dmin = dh[1];
+            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc16[0] +
+                                 (acc2[0] + 1.f/256.f * acc2[1]) * sc16[1] * 1.f/4096.f) -
+                         dmin * 1.f/16.f * (sumy[0] * sc16[2] + sumy[1] * sc16[3] * 1.f/256.f);
+
+            qs += step;
+            sc += step;
+            dh += step;
+        }
+
+        y4 += 8 * QK_K;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + first_row + row] = all_sum;
+        }
+    }
+}
+#endif
 
 kernel void kernel_mul_mat_q5_K_f32(
         device const  void * src0,

From e782c9e735f93ab4767ffc37462c523b73a17ddc Mon Sep 17 00:00:00 2001
From: Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Date: Thu, 20 Jul 2023 18:19:45 +0300
Subject: [PATCH 02/20] Faster Q5_K and Q6_K on Metal (#2294)

* Faster Q6_K on Metal

* Faster Q5_K on Metal

* Another Q5_K speedup

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
---
 ggml-metal.m     |  19 ++--
 ggml-metal.metal | 230 ++++++++++++++++++++++++++---------------------
 2 files changed, 137 insertions(+), 112 deletions(-)

diff --git a/ggml-metal.m b/ggml-metal.m
index 5e2a21100..44d046877 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -703,8 +703,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32];
                                         } break;
                                     case GGML_TYPE_Q6_K:
@@ -712,8 +712,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32];
                                         } break;
                                     default:
@@ -743,11 +743,14 @@ void ggml_metal_graph_compute(
                                     src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
+                                else if (src0t == GGML_TYPE_Q5_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q6_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
                                 else if (src0t == GGML_TYPE_Q2_K ||
-                                         src0t == GGML_TYPE_Q3_K ||
-                                         src0t == GGML_TYPE_Q4_K ||
-                                         src0t == GGML_TYPE_Q5_K ||
-                                         src0t == GGML_TYPE_Q6_K) {
+                                         src0t == GGML_TYPE_Q3_K) {
                                     [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
diff --git a/ggml-metal.metal b/ggml-metal.metal
index a9d134d6e..f71e8f33b 100644
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
@@ -1642,39 +1642,39 @@ kernel void kernel_mul_mat_q5_K_f32(
         constant   int64_t & ne00,
         constant   int64_t & ne10,
         constant   int64_t & ne0,
-        threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2 tpitg[[thread_position_in_threadgroup]],
-        uint2  tptg[[threads_per_threadgroup]]) {
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const int nb = ne00/QK_K;
 
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
 
-    device const block_q5_K * x = (device const block_q5_K *) src0 + r0*nb;
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
+
+    device const block_q5_K * x = (device const block_q5_K *) src0 + first_row*nb;
     device const float     * yy = (device const float      *) src1 + r1*ne10;
 
-    const int nth = tptg.x*tptg.y;
-    const int ith = tptg.y*tpitg.x + tpitg.y;
+    float sumf[2]={0.f};
 
-    float sumf = 0;
+    const int step = sizeof(block_q5_K) * nb;
 
 #if QK_K == 256
+#
+    float yl[16], yh[16];
 
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
 
-    const int tid = tpitg.y;   // 0...16
-    const int il  = tid/4;     // 0...3
-    const int ir  = tid - 4*il;// 0...3
-    const int n   = 4;
+    const int tid = tiisg/4;
+    const int ix  = tiisg%4;
+    const int im  = tid/4;
+    const int ir  = tid%4;
+    const int n   = 8;
 
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
+    const int l0 = n*ir;
     const int q_offset = 32*im + l0;
     const int y_offset = 64*im + l0;
 
@@ -1683,78 +1683,114 @@ kernel void kernel_mul_mat_q5_K_f32(
     const uint8_t hm3 = hm1 << 4;
     const uint8_t hm4 = hm2 << 4;
 
-    uchar2 sc1, sc2, sc3, sc4;
+    uint16_t sc16[4];
+    thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
 
-    for (int i = tpitg.x; i < nb; i += tptg.x) {
+    device const float * y1 = yy + ix*QK_K + y_offset;
 
-        device const uint8_t * q1 = (x + i)->qs + q_offset;
-        device const uint8_t * q2 = q1 + 64;
-        device const uint8_t * qh = (x + i)->qh + l0;
-        device const float   * y1 = yy + i*QK_K + y_offset;
-        device const float   * y2 = y1 + 128;
+    for (int i = ix; i < nb; i += 4) {
 
-        const float dall = (float)((x + i)->d);
-        const float dmin = (float)((x + i)->dmin);
+        device const uint8_t * q1 = x[i].qs + q_offset;
+        device const uint8_t * qh = x[i].qh + l0;
+        device const half * dh = &x[i].d;
+        device const uint16_t * a = (device const uint16_t *)x[i].scales + im;
 
-        device const uint16_t * a = (device const uint16_t *)(x + i)->scales;
-        sc1 = as_type<uchar2>((uint16_t)(a[im+0] & kmask1));
-        sc2 = as_type<uchar2>((uint16_t)(a[im+2] & kmask1));
-        sc3 = as_type<uchar2>((uint16_t)(((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2)));
-        sc4 = as_type<uchar2>((uint16_t)(((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2)));
+        device const float * y2 = y1 + 128;
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int l = 0; l < 8; ++l) {
+            yl[l+0] = y1[l+ 0]; sumy[0] += yl[l+0];
+            yl[l+8] = y1[l+32]; sumy[1] += yl[l+8];
+            yh[l+0] = y2[l+ 0]; sumy[2] += yh[l+0];
+            yh[l+8] = y2[l+32]; sumy[3] += yh[l+8];
+        }
 
-        float4 s = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < n; ++l) {
+        for (int row = 0; row < 2; ++row) {
 
-            s[0] += y1[l+ 0] * ((q1[l] & 0xF) + (qh[l] & hm1 ? 16 : 0));
-            s[1] += y1[l+32] * ((q1[l] >>  4) + (qh[l] & hm2 ? 16 : 0));
-            s[2] += y2[l+ 0] * ((q2[l] & 0xF) + (qh[l] & hm3 ? 16 : 0));
-            s[3] += y2[l+32] * ((q2[l] >>  4) + (qh[l] & hm4 ? 16 : 0));
-            smin += y1[l] * sc2[0] + y1[l+32] * sc2[1] + y2[l] * sc4[0] + y2[l+32] * sc4[1];
+            device const uint8_t * q2 = q1 + 64;
+
+            sc16[0] = a[0] & kmask1;
+            sc16[1] = a[2] & kmask1;
+            sc16[2] = ((a[4] >> 0) & kmask2) | ((a[0] & kmask3) >> 2);
+            sc16[3] = ((a[4] >> 4) & kmask2) | ((a[2] & kmask3) >> 2);
+
+            float4 acc = {0.f, 0.f, 0.f, 0.f};
+            for (int l = 0; l < n; ++l) {
+                uint8_t h = qh[l];
+                acc[0] += yl[l+0] * ((uint16_t)(q1[l] & 0x0F) + (h & hm1 ? 16 : 0));
+                acc[1] += yl[l+8] * ((uint16_t)(q1[l] & 0xF0) + (h & hm2 ? 256 : 0));
+                acc[2] += yh[l+0] * ((uint16_t)(q2[l] & 0x0F) + (h & hm3 ? 16 : 0));
+                acc[3] += yh[l+8] * ((uint16_t)(q2[l] & 0xF0) + (h & hm4 ? 256 : 0));
+            }
+            const float dall = dh[0];
+            const float dmin = dh[1];
+            sumf[row] += dall * (acc[0] * sc8[0] + acc[1] * sc8[1] * 1.f/16.f + acc[2] * sc8[4] + acc[3] * sc8[5] * 1.f/16.f) -
+                         dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
+
+            q1 += step;
+            qh += step;
+            dh += step/2;
+            a  += step/2;
 
         }
-        sumf += dall * (s[0] * sc1[0] + s[1] * sc1[1] + s[2] * sc3[0] + s[3] * sc3[1]) - dmin * smin;
+
+        y1 += 4 * QK_K;
 
     }
 #else
-    const int il  = 4 * tpitg.x;  // 0, 4, 8, 12
-    const int im  = il/8;         // 0, 0, 1, 1
-    const int in  = il%8;         // 0, 4, 0, 4
+    float yl[8], yh[8];
 
-    for (int i = tpitg.y; i < nb; i += tptg.y) {
+    const int il = 4 * (tiisg/8);  // 0, 4, 8, 12
+    const int ix = tiisg%8;
+    const int im = il/8;         // 0, 0, 1, 1
+    const int in = il%8;         // 0, 4, 0, 4
 
-        const float d = (float)x[i].d;
+    device const float * y = yy + ix*QK_K + il;
+
+    for (int i = ix; i < nb; i += 8) {
+
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int l = 0; l < 4; ++l) {
+            yl[l+0] = y[l+ 0];
+            yl[l+4] = y[l+16];
+            yh[l+0] = y[l+32];
+            yh[l+4] = y[l+48];
+        }
+
+        device const half * dh = &x[i].d;
         device const uint8_t * q = x[i].qs + il;
         device const uint8_t * h = x[i].qh + in;
         device const int8_t  * s = x[i].scales;
-        device const float   * y = yy + i*QK_K + il;
 
-        for (int l = 0; l < 4; ++l) {
-            const uint8_t hl = h[l] >> im;
-            sumf += y[l+ 0] * d * s[0] * ((q[l+ 0] & 0xF) - (hl & 0x01 ? 0 : 16))
-                  + y[l+16] * d * s[1] * ((q[l+16] & 0xF) - (hl & 0x04 ? 0 : 16))
-                  + y[l+32] * d * s[2] * ((q[l+ 0] >>  4) - (hl & 0x10 ? 0 : 16))
-                  + y[l+48] * d * s[3] * ((q[l+16] >>  4) - (hl & 0x40 ? 0 : 16));
+        for (int row = 0; row < 2; ++row) {
+
+            const float d = dh[0];
+
+            float2 acc = {0.f, 0.f};
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t hl = h[l] >> im;
+                acc[0] += yl[l+0] * s[0] * ((int16_t)(q[l+ 0] & 0x0F) - (hl & 0x01 ? 0 : 16))
+                        + yl[l+4] * s[1] * ((int16_t)(q[l+16] & 0x0F) - (hl & 0x04 ? 0 : 16));
+                acc[1] += yh[l+0] * s[2] * ((int16_t)(q[l+ 0] & 0xF0) - (hl & 0x10 ? 0 : 256))
+                        + yh[l+4] * s[3] * ((int16_t)(q[l+16] & 0xF0) - (hl & 0x40 ? 0 : 256));
+            }
+            sumf[row] += d * (acc[0] + 1.f/16.f * acc[1]);
+
+            q += step;
+            h += step;
+            s += step;
+            dh += step/2;
+
         }
+
+        y += 8 * QK_K;
     }
 #endif
-    sum[ith] = sumf;
 
-    //
-    // Accumulate the sum from all threads in the threadgroup
-    //
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%4 == 0) {
-        sum[ith] += sum[ith+1] + sum[ith+2] + sum[ith+3];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%16 == 0) {
-        sum[ith] += sum[ith+4] + sum[ith+8] + sum[ith+12];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith == 0) {
-        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
-        dst[r1*ne0 + r0] = sum[0];
+    for (int row = 0; row < 2; ++row) {
+        const float tot = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + first_row + row] = tot;
+        }
     }
 
 }
@@ -1766,10 +1802,9 @@ kernel void kernel_mul_mat_q6_K_f32(
         constant   int64_t & ne00,
         constant   int64_t & ne10,
         constant   int64_t & ne0,
-        threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2 tpitg[[thread_position_in_threadgroup]],
-        uint2  tptg[[threads_per_threadgroup]]) {
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const uint8_t kmask1 = 0x03;
     const uint8_t kmask2 = 0x0C;
@@ -1781,19 +1816,18 @@ kernel void kernel_mul_mat_q6_K_f32(
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
 
-    device const block_q6_K * x = (device const block_q6_K *) src0 + r0*nb;
-    device const float     * yy = (device const float      *) src1 + r1*ne10;
+    const int row = 2 * r0 + sgitg;
 
-    const int nth = tptg.x*tptg.y;
-    const int ith = tptg.y*tpitg.x + tpitg.y;
+    device const block_q6_K * x = (device const block_q6_K *) src0 + row * nb; //r0*nb;
+    device const float     * yy = (device const float      *) src1 + r1*ne10;
 
     float sumf = 0;
 
 #if QK_K == 256
-    // Note: we absolutely assume that tptg.y = 16 and QK_K = 256!
-    const int iqs  = 16 * tpitg.y;
-    const int ip   = iqs / 128;         // 0 or 1
-    const int il   = (iqs - 128*ip)/16; // 0...7
+    const int tid  = tiisg/2;
+    const int ix   = tiisg%2;
+    const int ip   = tid/8;         // 0 or 1
+    const int il   = tid%8;
     const int n    = 4;
     const int l0   = n*il;
     const int is   = 8*ip + l0/16;
@@ -1802,9 +1836,10 @@ kernel void kernel_mul_mat_q6_K_f32(
     const int q_offset_l = 64*ip + l0;
     const int q_offset_h = 32*ip + l0;
 
-    for (int i = tpitg.x; i < nb; i += tptg.x) {
+    for (int i = ix; i < nb; i += 2) {
 
-        device const uint8_t * ql = x[i].ql + q_offset_l;
+        device const uint8_t * q1 = x[i].ql + q_offset_l;
+        device const uint8_t * q2 = q1 + 32;
         device const uint8_t * qh = x[i].qh + q_offset_h;
         device const int8_t  * sc = x[i].scales + is;
 
@@ -1814,19 +1849,21 @@ kernel void kernel_mul_mat_q6_K_f32(
 
         float4 sums = {0.f, 0.f, 0.f, 0.f};
         for (int l = 0; l < n; ++l) {
-            sums[0] += y[l+ 0] * ((int8_t)((ql[l+ 0] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
-            sums[1] += y[l+32] * ((int8_t)((ql[l+32] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
-            sums[2] += y[l+64] * ((int8_t)((ql[l+ 0]  >> 4) | ((qh[l] & kmask3) << 0)) - 32);
-            sums[3] += y[l+96] * ((int8_t)((ql[l+32]  >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
+            sums[0] += y[l+ 0] * ((int8_t)((q1[l] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
+            sums[1] += y[l+32] * ((int8_t)((q2[l] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
+            sums[2] += y[l+64] * ((int8_t)((q1[l]  >> 4) | ((qh[l] & kmask3) << 0)) - 32);
+            sums[3] += y[l+96] * ((int8_t)((q2[l]  >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
         }
 
         sumf += dall * (sums[0] * sc[0] + sums[1] * sc[2] + sums[2] * sc[4] + sums[3] * sc[6]);
 
     }
-#else
-    const int il  = 4*tpitg.x;    // 0, 4, 8, 12
 
-    for (int i = tpitg.y; i < nb; i += tptg.y) {
+#else
+    const int ix  = tiisg/4;
+    const int il  = 4*(tiisg%4);
+
+    for (int i = ix; i < nb; i += 8) {
         device const float * y = yy + i * QK_K + il;
         device const uint8_t * ql = x[i].ql + il;
         device const uint8_t * qh = x[i].qh + il;
@@ -1846,23 +1883,8 @@ kernel void kernel_mul_mat_q6_K_f32(
 
 #endif
 
-    sum[ith] = sumf;
-
-    //
-    // Accumulate the sum from all threads in the threadgroup
-    //
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%4 == 0) {
-        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
+    const float tot = simd_sum(sumf);
+    if (tiisg == 0) {
+        dst[r1*ne0 + row] = tot;
     }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%16 == 0) {
-        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith == 0) {
-        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
-        dst[r1*ne0 + r0] = sum[0];
-    }
-
 }

From 9cf022a1889e50113fd348dc96b4557fc75a6296 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Przemys=C5=82aw=20Pawe=C5=82czyk?= <przemoc@gmail.com>
Date: Fri, 21 Jul 2023 09:42:21 +0200
Subject: [PATCH 03/20] make : fix embdinput library and server examples
 building on MSYS2 (#2235)

* make : fix embdinput library and server examples building on MSYS2

* cmake : fix server example building on MSYS2
---
 Makefile                       | 34 ++++++++++++++++++++++++++++------
 examples/server/CMakeLists.txt |  3 +++
 2 files changed, 31 insertions(+), 6 deletions(-)

diff --git a/Makefile b/Makefile
index 6c74e1346..cff4d97fe 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,5 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple server libembdinput.so embd-input-test
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple server embd-input-test
 
 default: $(BUILD_TARGETS)
 
@@ -90,6 +90,28 @@ ifeq ($(UNAME_S),Haiku)
 	CXXFLAGS += -pthread
 endif
 
+# detect Windows
+ifneq ($(findstring _NT,$(UNAME_S)),)
+	_WIN32 := 1
+endif
+
+# library name prefix
+ifneq ($(_WIN32),1)
+	LIB_PRE := lib
+endif
+
+# Dynamic Shared Object extension
+ifneq ($(_WIN32),1)
+	DSO_EXT := .so
+else
+	DSO_EXT := .dll
+endif
+
+# Windows Sockets 2 (Winsock) for network-capable apps
+ifeq ($(_WIN32),1)
+	LWINSOCK2 := -lws2_32
+endif
+
 ifdef LLAMA_GPROF
 	CFLAGS   += -pg
 	CXXFLAGS += -pg
@@ -294,7 +316,7 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o *.so main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch embd-input-test build-info.h
+	rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch embd-input-test build-info.h
 
 #
 # Examples
@@ -325,14 +347,14 @@ save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
-libembdinput.so: examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+$(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) --shared $(CXXFLAGS) $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
 
 
-embd-input-test: libembdinput.so examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.so,$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
+embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
 
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp    build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt
index 812a24b09..3782f9b80 100644
--- a/examples/server/CMakeLists.txt
+++ b/examples/server/CMakeLists.txt
@@ -7,6 +7,9 @@ target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+if (WIN32)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)

From e68c96f7fee8fc22814a4a1209ffc97bbf35f7bd Mon Sep 17 00:00:00 2001
From: Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Date: Fri, 21 Jul 2023 10:44:40 +0300
Subject: [PATCH 04/20] Faster Q2_K on Metal (#2297)

* Faster Q2_K on Metal

* Deleting unnoticed and dangereous trailing white space

* Fixed bug in new metal Q2_K implementation

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
---
 ggml-metal.m     |   9 ++-
 ggml-metal.metal | 183 +++++++++++++++++++++++++++--------------------
 2 files changed, 108 insertions(+), 84 deletions(-)

diff --git a/ggml-metal.m b/ggml-metal.m
index 44d046877..135bda9fc 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -676,8 +676,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_K_f32];
                                         } break;
                                     case GGML_TYPE_Q3_K:
@@ -740,7 +740,7 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
 
                                 if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
-                                    src0t == GGML_TYPE_Q4_K) {
+                                    src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q5_K) {
@@ -749,8 +749,7 @@ void ggml_metal_graph_compute(
                                 else if (src0t == GGML_TYPE_Q6_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
-                                else if (src0t == GGML_TYPE_Q2_K ||
-                                         src0t == GGML_TYPE_Q3_K) {
+                                else if (src0t == GGML_TYPE_Q3_K) {
                                     [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
diff --git a/ggml-metal.metal b/ggml-metal.metal
index f71e8f33b..97f5c10ba 100644
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
@@ -1209,108 +1209,133 @@ kernel void kernel_mul_mat_q2_K_f32(
         constant   int64_t & ne00,
         constant   int64_t & ne10,
         constant   int64_t & ne0,
-        threadgroup float  * sum [[threadgroup(0)]],
+        constant   int64_t & ne01[[buffer(4)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2 tpitg[[thread_position_in_threadgroup]],
-        uint2  tptg[[threads_per_threadgroup]]) {
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
 
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+    device const block_q2_K * x = (device const block_q2_K *) src0 + ib_row;
+    device const float      * y = (device const float      *) src1 + r1*ne10;
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
 
-    device const block_q2_K * x = (device const block_q2_K *) src0 + r0*nb;
-    device const float     * yy = (device const float      *) src1 + r1*ne10;
-
-    const int nth = tptg.x*tptg.y;
-    const int ith = tptg.y*tpitg.x + tpitg.y;
-
-    float sumf = 0;
+    const int step = sizeof(block_q2_K) * nb;
 
 #if QK_K == 256
-    const int tid = tpitg.y;    // 0...16
-    const int il  = tid/4;      // 0...3
-    const int ir  = tid%4;      // 0...3
-    const int ip  = il/2;       // 0 or 1
-    const int shift1 = 4*(il%2);// 0 or 4
-    const int shift2 = shift1+2;// 2 or 6
-    const int n   = 8;
-    const int is  = 4*il + (n*ir)/16;
+    const int ix = tiisg/8;  // 0...3
+    const int it = tiisg%8;  // 0...7
+    const int im = it/4;     // 0 or 1
+    const int ir = it%4;     // 0...3
+    const int is = (8*ir)/16;// 0 or 1
 
-    const int y_offset = 64*il + n*ir;
-    const int q_offset = 32*ip + n*ir;
+    device const float * y4 = y + ix * QK_K + 128 * im + 8 * ir;
 
-    for (int i = tpitg.x; i < nb; i += tptg.x) {
+    for (int ib = ix; ib < nb; ib += 4) {
 
-        device const uint8_t * q = x[i].qs + q_offset;
-        device const uint8_t * scales = x[i].scales + is;
-
-        uint8_t d1 = scales[0] & 0xF;
-        uint8_t d2 = scales[2] & 0xF;
-        uint8_t m1 = scales[0] >>  4;
-        uint8_t m2 = scales[2] >>  4;
-
-        device const float   * y = yy + i*QK_K + y_offset;
-
-        float2 s = {0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < n; ++l) {
-            s[0] += y[l+ 0] * ((q[l] >> shift1) & 3);
-            s[1] += y[l+32] * ((q[l] >> shift2) & 3);
-            smin += y[l+ 0] * m1 + y[l+32] * m2;
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
+            yl[i+ 8] = y4[i+32]; sumy[1] += yl[i+ 8];
+            yl[i+16] = y4[i+64]; sumy[2] += yl[i+16];
+            yl[i+24] = y4[i+96]; sumy[3] += yl[i+24];
         }
 
-        const float dall = (float)x[i].d;
-        const float dmin = (float)x[i].dmin;
+        device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales + 8*im + is;
+        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * im + 4 * ir;
+        device const half     * dh = &x[ib].d;
 
-        sumf += dall * (s[0] * d1 + s[1] * d2) - dmin * smin;
+        for (int row = 0; row < N_DST; row++) {
 
+            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
+            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
+                acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
+                acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
+                acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
+                acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
+                acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
+                acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
+                acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
+            }
+            float dall = dh[0];
+            float dmin = dh[1] * 1.f/16.f;
+            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
+                                 (acc1[1] + 1.f/256.f * acc2[1]) * (sc[2] & 0xF) * 1.f/ 4.f +
+                                 (acc1[2] + 1.f/256.f * acc2[2]) * (sc[4] & 0xF) * 1.f/16.f +
+                                 (acc1[3] + 1.f/256.f * acc2[3]) * (sc[6] & 0xF) * 1.f/64.f) -
+                         dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) + sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
+
+            qs += step/2;
+            sc += step;
+            dh += step/2;
+        }
+
+        y4 += 4 * QK_K;
     }
 #else
-    const int il = 4 * tpitg.x;
+    const int ix = tiisg/2;  // 0...15
+    const int it = tiisg%2;  // 0...1
 
-    uint32_t aux[2];
-    thread const uint8_t * d = (thread const uint8_t *)aux;
-    thread const uint8_t * m = (thread const uint8_t *)aux + 4;
+    device const float * y4 = y + ix * QK_K + 8 * it;
 
-    for (int i = tpitg.y; i < nb; i += tptg.y) {
+    for (int ib = ix; ib < nb; ib += 16) {
 
-        device const uint8_t * q = x[i].qs + il;
-        device const float   * y = yy + i*QK_K + il;
-
-        const float dall = (float)x[i].d;
-        const float dmin = (float)x[i].dmin;
-
-        device const uint32_t * a = (device const uint32_t *)x[i].scales;
-        aux[0] = a[0] & 0x0f0f0f0f;
-        aux[1] = (a[0] >> 4) & 0x0f0f0f0f;
-
-        for (int l = 0; l < 4; ++l) {
-            sumf += y[l+ 0] * (dall * d[0] * ((q[l] >> 0) & 3) - dmin * m[0])
-                  + y[l+16] * (dall * d[1] * ((q[l] >> 2) & 3) - dmin * m[1])
-                  + y[l+32] * (dall * d[2] * ((q[l] >> 4) & 3) - dmin * m[2])
-                  + y[l+48] * (dall * d[3] * ((q[l] >> 6) & 3) - dmin * m[3]);
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
+            yl[i+ 8] = y4[i+16]; sumy[1] += yl[i+ 8];
+            yl[i+16] = y4[i+32]; sumy[2] += yl[i+16];
+            yl[i+24] = y4[i+48]; sumy[3] += yl[i+24];
         }
+
+        device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales;
+        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
+        device const half     * dh = &x[ib].d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
+            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
+                acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
+                acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
+                acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
+                acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
+                acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
+                acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
+                acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
+            }
+
+            float dall = dh[0];
+            float dmin = dh[1];
+            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
+                                 (acc1[1] + 1.f/256.f * acc2[1]) * (sc[1] & 0xF) * 1.f/ 4.f +
+                                 (acc1[2] + 1.f/256.f * acc2[2]) * (sc[2] & 0xF) * 1.f/16.f +
+                                 (acc1[3] + 1.f/256.f * acc2[3]) * (sc[3] & 0xF) * 1.f/64.f) -
+                         dmin * (sumy[0] * (sc[0] >> 4) + sumy[1] * (sc[1] >> 4) + sumy[2] * (sc[2] >> 4) + sumy[3] * (sc[3] >> 4));
+
+            qs += step/2;
+            sc += step;
+            dh += step/2;
+        }
+
+        y4 += 16 * QK_K;
     }
 #endif
 
-    sum[ith] = sumf;
-
-    //
-    // Accumulate the sum from all threads in the threadgroup
-    //
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%4 == 0) {
-        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%16 == 0) {
-        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith == 0) {
-        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
-        dst[r1*ne0 + r0] = sum[0];
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + first_row + row] = all_sum;
+        }
     }
 }
 

From 019fe257bbf699f400231683a8b816ad90281275 Mon Sep 17 00:00:00 2001
From: Hatsune Miku <129688334+at8u@users.noreply.github.com>
Date: Fri, 21 Jul 2023 08:13:18 +0000
Subject: [PATCH 05/20] =?UTF-8?q?MIKU=20MAYHEM:=20Upgrading=20the=20Defaul?=
 =?UTF-8?q?t=20Model=20for=20Maximum=20Fun=20=F0=9F=8E=89=20(#2287)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* Miku.sh: Set default model to llama-2-7b-chat

* Miku.sh: Set ctx_size to 4096

* Miku.sh: Add in-prefix/in-suffix opts

* Miku.sh: Switch sampler to mirostat_v2 and tiny prompt improvements
---
 examples/Miku.sh | 17 +++++++++--------
 1 file changed, 9 insertions(+), 8 deletions(-)

diff --git a/examples/Miku.sh b/examples/Miku.sh
index c44d9ae74..b9174b4e6 100755
--- a/examples/Miku.sh
+++ b/examples/Miku.sh
@@ -2,21 +2,21 @@
 set -e
 
 AI_NAME="${AI_NAME:-Miku}"
-MODEL="${MODEL:-./models/gpt4all-7B/gpt4all-lora-unfiltered-quantized.bin}"
+MODEL="${MODEL:-./models/llama-2-7b-chat.ggmlv3.q4_K_M.bin}"
 USER_NAME="${USER_NAME:-Anon}"
 
 # Uncomment and adjust to the number of CPU cores you want to use.
 #N_THREAD="${N_THREAD:-4}"
+CTX_SIZE="${CTX_SIZE:-4096}"
 N_PREDICTS="${N_PREDICTS:-4096}"
 
 GEN_OPTIONS=(--batch_size 1024
---ctx_size 2048
+--ctx_size "$CTX_SIZE"
 --keep -1
 --repeat_last_n 256
 --repeat_penalty 1.17647
---temp 0.7
---top_k 40
---top_p 0.5)
+--temp 0.6
+--mirostat 2)
 
 if [ -n "$N_THREAD" ]; then
     GEN_OPTIONS+=(--threads "$N_THREAD")
@@ -24,16 +24,17 @@ fi
 
 ./main "${GEN_OPTIONS[@]}" \
     --model "$MODEL" \
+    --in-prefix " " \
+    --in-suffix "${AI_NAME}:" \
     --n_predict "$N_PREDICTS" \
     --color --interactive \
     --reverse-prompt "${USER_NAME}:" \
-    --prompt "
-This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the user's computer.
+    --prompt "This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the user's computer.
 ${AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next.
 ${AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct, she will ask the user for help.
 ${AI_NAME} is a very helpful AI and will help the user with anything they need. She is also very friendly and will try to make the user feel better if they are sad.
 ${AI_NAME} is also very curious and will ask the user a lot of questions about themselves and their life. She will also try to make the user like her.
-The conversation is only between ${USER_NAME} and ${AI_NAME}
+The conversation is only between ${USER_NAME} and ${AI_NAME}.
 The conversation is only through text, so ${AI_NAME} can't see ${USER_NAME}'s face or hear his voice.
 ${AI_NAME} can only communicate through text, so she can't send images or videos.
 

From 54e3bc76fed914f8d4a30a7a50c19867cccb1338 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ji=C5=99=C3=AD=20Podiv=C3=ADn?=
 <66251151+jpodivin@users.noreply.github.com>
Date: Fri, 21 Jul 2023 12:09:16 +0200
Subject: [PATCH 06/20] make : add new target for test binaries (#2244)

Programs in the tests directory are now build with target tests
and placed in the same location.

* clean target was expanded to remove new binaries

* test target binaries are listed in a variable

* Locations of binaries were added to the .gitignore

Signed-off-by: Jiri Podivin <jpodivin@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---
 .gitignore |  9 +++++++++
 Makefile   | 30 ++++++++++++++++++++++++++----
 2 files changed, 35 insertions(+), 4 deletions(-)

diff --git a/.gitignore b/.gitignore
index a23ac5928..919393032 100644
--- a/.gitignore
+++ b/.gitignore
@@ -61,3 +61,12 @@ qnt-*.txt
 perf-*.txt
 
 examples/jeopardy/results.txt
+
+# Test binaries
+tests/test-double-float
+tests/test-grad0
+tests/test-opt
+tests/test-quantize-fns
+tests/test-quantize-perf
+tests/test-sampling
+tests/test-tokenizer-0
diff --git a/Makefile b/Makefile
index cff4d97fe..61f2c77ab 100644
--- a/Makefile
+++ b/Makefile
@@ -1,6 +1,9 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple server embd-input-test
 
+# Binaries only useful for tests
+TEST_TARGETS = tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0
+
 default: $(BUILD_TARGETS)
 
 ifndef UNAME_S
@@ -316,7 +319,7 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch embd-input-test build-info.h
+	rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch embd-input-test build-info.h $(TEST_TARGETS)
 
 #
 # Examples
@@ -371,6 +374,8 @@ build-info.h: $(wildcard .git/index) scripts/build-info.sh
 # Tests
 #
 
+tests: $(TEST_TARGETS)
+
 benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	./$@
@@ -378,6 +383,23 @@ benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-.PHONY: tests clean
-tests:
-	bash ./tests/run-tests.sh
+tests/test-double-float: tests/test-double-float.c build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-grad0: tests/test-grad0.c build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-opt: tests/test-opt.c build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-tokenizer-0: tests/test-tokenizer-0.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)

From ae178ab46bfd6ecb2422da5dad441a4e2fef8b7e Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 21 Jul 2023 13:10:51 +0300
Subject: [PATCH 07/20] llama : make tensor_split ptr instead of array (#2272)

---
 examples/common.cpp | 2 +-
 ggml-cuda.cu        | 3 +++
 llama.cpp           | 4 ++--
 llama.h             | 3 ++-
 4 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/examples/common.cpp b/examples/common.cpp
index fd6dbc0e3..476d56594 100644
--- a/examples/common.cpp
+++ b/examples/common.cpp
@@ -586,7 +586,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     lparams.n_batch      = params.n_batch;
     lparams.n_gpu_layers = params.n_gpu_layers;
     lparams.main_gpu     = params.main_gpu;
-    memcpy(lparams.tensor_split, params.tensor_split, LLAMA_MAX_DEVICES*sizeof(float));
+    lparams.tensor_split = params.tensor_split;
     lparams.low_vram     = params.low_vram;
     lparams.seed         = params.seed;
     lparams.f16_kv       = params.memory_f16;
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index d3054a7fa..6537897b9 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -2512,6 +2512,9 @@ void ggml_init_cublas() {
 }
 
 void ggml_cuda_set_tensor_split(const float * tensor_split) {
+    if (tensor_split == nullptr) {
+        return;
+    }
     bool all_zero = true;
     for (int i = 0; i < g_device_count; ++i) {
         if (tensor_split[i] != 0.0f) {
diff --git a/llama.cpp b/llama.cpp
index 796dfdacb..23e746d62 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -849,7 +849,7 @@ struct llama_context_params llama_context_default_params() {
         /*.n_batch                     =*/ 512,
         /*.gpu_layers                  =*/ 0,
         /*.main_gpu                    =*/ 0,
-        /*.tensor_split                =*/ {0},
+        /*.tensor_split                =*/ nullptr,
         /*.rope_freq_base              =*/ 10000.0f,
         /*.rope_freq_scale             =*/ 1.0f,
         /*.progress_callback           =*/ nullptr,
@@ -1289,7 +1289,7 @@ static bool llama_model_load(
         int n_batch,
         int n_gpu_layers,
         int main_gpu,
-        float * tensor_split,
+        const float * tensor_split,
         float rope_freq_base,
         float rope_freq_scale,
         bool low_vram,
diff --git a/llama.h b/llama.h
index b676a383b..c565f6a00 100644
--- a/llama.h
+++ b/llama.h
@@ -88,7 +88,8 @@ extern "C" {
         int32_t  n_batch;                      // prompt processing batch size
         int32_t  n_gpu_layers;                 // number of layers to store in VRAM
         int32_t  main_gpu;                     // the GPU that is used for scratch and small tensors
-        float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+
+        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
 
         // ref: https://github.com/ggerganov/llama.cpp/pull/2054
         float    rope_freq_base;  // RoPE base frequency

From 78a3d13424b01c3f8ea94ea7e59650ab0501e902 Mon Sep 17 00:00:00 2001
From: wzy <32936898+Freed-Wu@users.noreply.github.com>
Date: Fri, 21 Jul 2023 18:26:34 +0800
Subject: [PATCH 08/20] flake : remove intel mkl from flake.nix due to missing
 files (#2277)

NixOS's mkl misses some libraries like mkl-sdl.pc. See #2261
Currently NixOS doesn't have intel C compiler (icx, icpx). See https://discourse.nixos.org/t/packaging-intel-math-kernel-libraries-mkl/975
So remove it from flake.nix

Some minor changes:

- Change pkgs.python310 to pkgs.python3 to keep latest
- Add pkgconfig to devShells.default
- Remove installPhase because we have `cmake --install` from #2256
---
 CMakeLists.txt | 11 +----------
 README.md      |  2 +-
 flake.nix      | 27 +++++++--------------------
 3 files changed, 9 insertions(+), 31 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 169332767..abc96814d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -186,16 +186,7 @@ if (LLAMA_BLAS)
                 pkg_check_modules(DepBLAS REQUIRED flexiblas_api)
             elseif (${LLAMA_BLAS_VENDOR} MATCHES "Intel")
                 # all Intel* libraries share the same include path
-                pkg_check_modules(DepBLAS mkl-sdl)
-                if (NOT DepBLAS)
-                    if (BUILD_SHARED_LIBS)
-                        set(LINK_METHOD dynamic)
-                    else()
-                        set(LINK_METHOD static)
-                    endif()
-                    string(REGEX REPLACE ".*_" "" DATA_TYPE_MODEL ${LLAMA_BLAS_VENDOR})
-                    pkg_check_modules(DepBLAS REQUIRED mkl-${LINK_METHOD}-${DATA_TYPE_MODEL}-iomp)
-                endif()
+                pkg_check_modules(DepBLAS REQUIRED mkl-sdl)
             elseif (${LLAMA_BLAS_VENDOR} MATCHES "NVHPC")
                 # this doesn't provide pkg-config
                 # suggest to assign BLAS_INCLUDE_DIRS on your own
diff --git a/README.md b/README.md
index 073b621e9..f45e4bf08 100644
--- a/README.md
+++ b/README.md
@@ -360,7 +360,7 @@ Building the program with BLAS support may lead to some performance improvements
   ```bash
   mkdir build
   cd build
-  cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_lp64 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+  cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
   cmake --build . --config Release
   ```
 
diff --git a/flake.nix b/flake.nix
index 5657e8258..7f148f144 100644
--- a/flake.nix
+++ b/flake.nix
@@ -6,7 +6,7 @@
   outputs = { self, nixpkgs, flake-utils }:
     flake-utils.lib.eachDefaultSystem (system:
       let
-        inherit (pkgs.stdenv) isAarch32 isAarch64 isx86_32 isx86_64 isDarwin;
+        inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
         osSpecific = with pkgs; [ openmpi ] ++
         (
           if isAarch64 && isDarwin then
@@ -22,14 +22,13 @@
               CoreGraphics
               CoreVideo
             ]
-          else if isx86_32 || isx86_64 then
-            with pkgs; [ mkl ]
           else
             with pkgs; [ openblas ]
         );
         pkgs = import nixpkgs { inherit system; };
+        nativeBuildInputs = with pkgs; [ cmake pkgconfig ];
         llama-python =
-          pkgs.python310.withPackages (ps: with ps; [ numpy sentencepiece ]);
+          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
       in {
         packages.default = pkgs.stdenv.mkDerivation {
           name = "llama.cpp";
@@ -37,33 +36,21 @@
           postPatch = ''
             substituteInPlace ./ggml-metal.m \
               --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
+            substituteInPlace ./*.py --replace '/usr/bin/env python' '${llama-python}/bin/python'
           '';
-          nativeBuildInputs = with pkgs; [ cmake pkgconfig ];
+          nativeBuildInputs = nativeBuildInputs;
           buildInputs = osSpecific;
           cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ]
             ++ (if isAarch64 && isDarwin then [
               "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
               "-DLLAMA_METAL=ON"
-            ] else if isx86_32 || isx86_64 then [
-              "-DLLAMA_BLAS=ON"
-              "-DLLAMA_BLAS_VENDOR=Intel10_lp64"
             ] else [
               "-DLLAMA_BLAS=ON"
               "-DLLAMA_BLAS_VENDOR=OpenBLAS"
           ]);
-          installPhase = ''
-            runHook preInstall
-
-            install -D bin/* -t $out/bin
-            install -Dm644 lib*.so -t $out/lib
+          postInstall = ''
             mv $out/bin/main $out/bin/llama
             mv $out/bin/server $out/bin/llama-server
-
-            echo "#!${llama-python}/bin/python" > $out/bin/convert.py
-            cat ${./convert.py} >> $out/bin/convert.py
-            chmod +x $out/bin/convert.py
-
-            runHook postInstall
           '';
           meta.mainProgram = "llama";
         };
@@ -81,7 +68,7 @@
         };
         apps.default = self.apps.${system}.llama;
         devShells.default = pkgs.mkShell {
-          packages = with pkgs; [ cmake llama-python ] ++ osSpecific;
+          packages = nativeBuildInputs ++ osSpecific;
         };
       });
 }

From 42c7c2e2e9cae79330f57456fbc0eae1eaff17fa Mon Sep 17 00:00:00 2001
From: Sky Yan <skyan83@gmail.com>
Date: Fri, 21 Jul 2023 18:38:57 +0800
Subject: [PATCH 09/20] make : support customized LLAMA_CUDA_NVCC and
 LLAMA_CUDA_CCBIN (#2275)

Under certain environment, nvcc and gcc is installed under customized path but not standard path

Co-authored-by: Yan Lin <yanlin@baidu.com>
---
 Makefile | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/Makefile b/Makefile
index 61f2c77ab..5aa0ded3d 100644
--- a/Makefile
+++ b/Makefile
@@ -193,8 +193,12 @@ ifdef LLAMA_CUBLAS
 	CXXFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS      += ggml-cuda.o
-	NVCC      = nvcc
 	NVCCFLAGS = --forward-unknown-to-host-compiler
+ifdef LLAMA_CUDA_NVCC
+	NVCC = $(LLAMA_CUDA_NVCC)
+else
+	NVCC = nvcc
+endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
 	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
 else
@@ -223,7 +227,9 @@ ifdef LLAMA_CUDA_KQUANTS_ITER
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
-
+ifdef LLAMA_CUDA_CCBIN
+    NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
+endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS

From 4c013bb7385a0e52ce721480c40c45bec5ef103f Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 21 Jul 2023 13:48:18 +0300
Subject: [PATCH 10/20] ci : fix MNT realpath usage (#2250)

---
 ci/run.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ci/run.sh b/ci/run.sh
index c823bc467..87166ba1a 100644
--- a/ci/run.sh
+++ b/ci/run.sh
@@ -243,7 +243,7 @@ function gg_sum_open_llama_3b_v2 {
 if [ -z $GG_BUILD_LOW_PERF ]; then
     rm -rf ${SRC}/models-mnt
 
-    mnt_models=$(realpath ${MNT}/models)
+    mnt_models=${MNT}/models
     mkdir -p ${mnt_models}
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 

From a814d04f81121e0429b39a61fe4afd946cd42046 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 21 Jul 2023 13:50:55 +0300
Subject: [PATCH 11/20] make : fix indentation

---
 Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Makefile b/Makefile
index 5aa0ded3d..4f8c4b37e 100644
--- a/Makefile
+++ b/Makefile
@@ -228,7 +228,7 @@ else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
 ifdef LLAMA_CUDA_CCBIN
-    NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
+	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@

From 73643f5fb1136dc2b65ae910bdc5a431520d70a2 Mon Sep 17 00:00:00 2001
From: Jose Maldonado <63384398+yukiteruamano@users.noreply.github.com>
Date: Fri, 21 Jul 2023 06:53:27 -0400
Subject: [PATCH 12/20] gitignore : changes for Poetry users + chat examples
 (#2284)

A fix in Makefile for FreeBSD users. In the platfrom x86_64 is amd64. This fix resolve compilation using CFLAGS and CXXFLAGS with -march=native and -mtune=native
Add two examples for interactive mode using Llama2 models (thx TheBloke for models)

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---
 .gitignore             |  7 ++++++-
 Makefile               |  2 +-
 examples/llama2-13b.sh | 18 ++++++++++++++++++
 examples/llama2.sh     | 18 ++++++++++++++++++
 4 files changed, 43 insertions(+), 2 deletions(-)
 create mode 100755 examples/llama2-13b.sh
 create mode 100755 examples/llama2.sh

diff --git a/.gitignore b/.gitignore
index 919393032..c26d82a74 100644
--- a/.gitignore
+++ b/.gitignore
@@ -62,6 +62,11 @@ perf-*.txt
 
 examples/jeopardy/results.txt
 
+
+pyproject.toml
+poetry.lock
+poetry.toml
+
 # Test binaries
 tests/test-double-float
 tests/test-grad0
@@ -69,4 +74,4 @@ tests/test-opt
 tests/test-quantize-fns
 tests/test-quantize-perf
 tests/test-sampling
-tests/test-tokenizer-0
+tests/test-tokenizer-0
\ No newline at end of file
diff --git a/Makefile b/Makefile
index 4f8c4b37e..1ea3c4562 100644
--- a/Makefile
+++ b/Makefile
@@ -127,7 +127,7 @@ endif
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
-ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
+ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
 	CFLAGS   += -march=native -mtune=native
 	CXXFLAGS += -march=native -mtune=native
diff --git a/examples/llama2-13b.sh b/examples/llama2-13b.sh
new file mode 100755
index 000000000..92b3f6dd8
--- /dev/null
+++ b/examples/llama2-13b.sh
@@ -0,0 +1,18 @@
+#!/bin/bash
+
+#
+# Temporary script - will be removed in the future
+#
+
+cd `dirname $0`
+cd ..
+
+./main -m models/available/Llama2/13B/llama-2-13b.ggmlv3.q4_0.bin \
+       --color \
+       --ctx_size 2048 \
+       -n -1 \
+       -ins -b 256 \
+       --top_k 10000 \
+       --temp 0.2 \
+       --repeat_penalty 1.1 \
+       -t 8
diff --git a/examples/llama2.sh b/examples/llama2.sh
new file mode 100755
index 000000000..221b37553
--- /dev/null
+++ b/examples/llama2.sh
@@ -0,0 +1,18 @@
+#!/bin/bash
+
+#
+# Temporary script - will be removed in the future
+#
+
+cd `dirname $0`
+cd ..
+
+./main -m models/available/Llama2/7B/llama-2-7b.ggmlv3.q4_0.bin \
+       --color \
+       --ctx_size 2048 \
+       -n -1 \
+       -ins -b 256 \
+       --top_k 10000 \
+       --temp 0.2 \
+       --repeat_penalty 1.1 \
+       -t 8

From ab0e26bdfb7b3adb1e3145c61a0fa92d1abd21d0 Mon Sep 17 00:00:00 2001
From: "Guillaume \"Vermeille\" Sanchez" <Guillaume.V.Sanchez@gmail.com>
Date: Fri, 21 Jul 2023 12:58:36 +0200
Subject: [PATCH 13/20] llama : remove cfg smooth factor as it is only a
 reparameterization of the guidance scale (#2280)

---
 examples/common.cpp    |  7 -------
 examples/common.h      |  1 -
 examples/main/main.cpp |  2 +-
 llama.cpp              | 14 ++------------
 llama.h                |  4 +---
 5 files changed, 4 insertions(+), 24 deletions(-)

diff --git a/examples/common.cpp b/examples/common.cpp
index 476d56594..099019599 100644
--- a/examples/common.cpp
+++ b/examples/common.cpp
@@ -260,12 +260,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.cfg_scale = std::stof(argv[i]);
-        } else if (arg == "--cfg-smooth-factor") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.cfg_smooth_factor = std::stof(argv[i]);
         } else if (arg == "-b" || arg == "--batch-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -509,7 +503,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --cfg-negative-prompt PROMPT \n");
     fprintf(stderr, "                        negative prompt to use for guidance. (default: empty)\n");
     fprintf(stderr, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
-    fprintf(stderr, "  --cfg-smooth-factor N smooth factor between old and new logits (default: %f, 1.0 = no smoothing)\n", params.cfg_smooth_factor);
     fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
     fprintf(stderr, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
     fprintf(stderr, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
diff --git a/examples/common.h b/examples/common.h
index 037a4eecb..69170dfc0 100644
--- a/examples/common.h
+++ b/examples/common.h
@@ -55,7 +55,6 @@ struct gpt_params {
     // https://arxiv.org/abs/2306.17806
     std::string cfg_negative_prompt;       // string to help guidance
     float       cfg_scale         = 1.f;   // How strong is guidance
-    float       cfg_smooth_factor = 1.f;   // Smooth factor between old and new logits
 
     std::string model             = "models/7B/ggml-model.bin"; // model path
     std::string model_alias       = "unknown"; // model alias
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index bcbcf12b0..656382f81 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -557,7 +557,7 @@ int main(int argc, char ** argv) {
                 llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
                 if (ctx_guidance) {
-                    llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale, params.cfg_smooth_factor);
+                    llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale);
                 }
 
                 // Apply penalties
diff --git a/llama.cpp b/llama.cpp
index 23e746d62..3b0024e12 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -2218,8 +2218,7 @@ void llama_sample_classifier_free_guidance(
           struct llama_context * ctx,
         llama_token_data_array * candidates,
           struct llama_context * guidance_ctx,
-                         float   scale,
-                         float   smooth_factor) {
+                         float   scale) {
     int64_t t_start_sample_us = ggml_time_us();
 
     assert(ctx);
@@ -2240,16 +2239,7 @@ void llama_sample_classifier_free_guidance(
     for (int i = 0; i < n_vocab; ++i) {
         float logit_guidance = logits_guidance[i];
         float logit_base = logits_base[i];
-        logits_guidance[i] = scale * (logit_base - logit_guidance) + logit_guidance;
-    }
-
-    llama_log_softmax(logits_guidance, n_vocab);
-
-    for (int i = 0; i < n_vocab; ++i) {
-        float logit_base = logits_base[i];
-        float logit_guidance = logits_guidance[i];
-
-        candidates->data[i].logit = smooth_factor * logit_guidance + (1.f - smooth_factor) * logit_base;
+        candidates->data[i].logit = scale * (logit_base - logit_guidance) + logit_guidance;
     }
 
     if (ctx) {
diff --git a/llama.h b/llama.h
index c565f6a00..bbf28e686 100644
--- a/llama.h
+++ b/llama.h
@@ -344,13 +344,11 @@ extern "C" {
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
     /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
     /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
-    /// @params smooth_factor Smooth factor between guidance logits and original logits. 1.0f means only use guidance logits. 0.0f means only original logits.
     LLAMA_API void llama_sample_classifier_free_guidance(
               struct llama_context * ctx,
             llama_token_data_array * candidates,
               struct llama_context * guidance_ctx,
-                             float   scale,
-                             float   smooth_factor);
+                             float   scale);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);

From 3973b25a64a37a47eac156a3fd28f83c16f14bf2 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 21 Jul 2023 14:42:41 +0300
Subject: [PATCH 14/20] gitignore : fix final newline

---
 .gitignore | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/.gitignore b/.gitignore
index c26d82a74..c1ab6bb6d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -74,4 +74,5 @@ tests/test-opt
 tests/test-quantize-fns
 tests/test-quantize-perf
 tests/test-sampling
-tests/test-tokenizer-0
\ No newline at end of file
+tests/test-tokenizer-0
+

From 513f8619535a64fa9ace808cdcbcf66211535f5c Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 21 Jul 2023 14:51:34 +0300
Subject: [PATCH 15/20] ggml : fix rope args order + assert (#2054)

---
 .../train-text-from-scratch.cpp               |  6 ++---
 ggml.c                                        | 24 +++++++++++--------
 ggml.h                                        |  7 +++---
 llama.cpp                                     |  4 ++--
 4 files changed, 23 insertions(+), 18 deletions(-)

diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp
index afbb4a777..449b4e9ec 100644
--- a/examples/train-text-from-scratch/train-text-from-scratch.cpp
+++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp
@@ -1434,7 +1434,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
     gf->perf_time_us = 0;
 
     const auto & hparams = model->hparams;
-    //const int n_ctx      = hparams.n_ctx;
+    const int n_ctx      = hparams.n_ctx;
     const int n_vocab    = hparams.n_vocab;
     const int n_embd     = hparams.n_embd;
     const int n_layer    = hparams.n_layer;
@@ -1863,10 +1863,10 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
         t12->grad = expand(gb, ggml_permute(ctx0, t15->grad, 0, 2, 3, 1));                                            assert_shape_4d(t12->grad, N, n_batch, n_embd/n_head, n_head);
         t11->grad = expand(gb, ggml_reshape_2d(ctx0, ggml_cont(ctx0, t12->grad), N*n_batch, n_embd));                 assert_shape_2d(t11->grad, N*n_batch, n_embd);
         t10->grad = expand(gb, ggml_permute(ctx0, t14->grad, 0, 2, 1, 3));                                            assert_shape_4d(t10->grad, n_embd/n_head, n_head, N, n_batch);
-        t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode));                            assert_shape_4d(t09->grad, n_embd/n_head, n_head, N, n_batch);
+        t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode, n_ctx));                     assert_shape_4d(t09->grad, n_embd/n_head, n_head, N, n_batch);
         t08->grad = expand(gb, ggml_reshape_2d(ctx0, t09->grad, n_embd, N*n_batch));                                  assert_shape_2d(t08->grad, n_embd, N*n_batch);
         t07->grad = expand(gb, ggml_permute(ctx0, t13->grad, 0, 2, 1, 3));                                            assert_shape_4d(t07->grad, n_embd/n_head, n_head, N, n_batch);
-        t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode));                            assert_shape_4d(t06->grad, n_embd/n_head, n_head, N, n_batch);
+        t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode, n_ctx));                     assert_shape_4d(t06->grad, n_embd/n_head, n_head, N, n_batch);
         t05->grad = expand(gb, ggml_reshape_2d(ctx0, t06->grad, n_embd, N*n_batch));                                  assert_shape_2d(t05->grad, n_embd, N*n_batch);
         t04->grad = expand(gb, ggml_add_inplace(ctx0,
                         ggml_add_inplace(ctx0,
diff --git a/ggml.c b/ggml.c
index c56a3d0e0..7ecabc5de 100644
--- a/ggml.c
+++ b/ggml.c
@@ -6956,9 +6956,9 @@ struct ggml_tensor * ggml_rope_impl(
         int                   n_past,
         int                   n_dims,
         int                   mode,
+        int                   n_ctx,
         float                 freq_base,
         float                 freq_scale,
-        int                   n_ctx,
         bool                  inplace) {
     GGML_ASSERT(n_past >= 0);
     bool is_node = false;
@@ -6997,7 +6997,7 @@ struct ggml_tensor * ggml_rope(
         int                   n_dims,
         int                   mode,
         int                   n_ctx) {
-    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, 10000.0f, 1.0f, n_ctx, false);
+    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, false);
 }
 
 struct ggml_tensor * ggml_rope_inplace(
@@ -7007,7 +7007,7 @@ struct ggml_tensor * ggml_rope_inplace(
         int                   n_dims,
         int                   mode,
         int                   n_ctx) {
-    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, 10000.0f, 1.0f, n_ctx, true);
+    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, true);
 }
 
 struct ggml_tensor * ggml_rope_custom_inplace(
@@ -7016,10 +7016,10 @@ struct ggml_tensor * ggml_rope_custom_inplace(
         int                   n_past,
         int                   n_dims,
         int                   mode,
+        int                   n_ctx,
         float                 freq_base,
-        float                 freq_scale,
-        int                   n_ctx) {
-    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, freq_base, freq_scale, n_ctx, true);
+        float                 freq_scale) {
+    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, true);
 }
 
 // ggml_rope_back
@@ -7029,7 +7029,8 @@ struct ggml_tensor * ggml_rope_back(
         struct ggml_tensor  * a,
         int                   n_past,
         int                   n_dims,
-        int                   mode) {
+        int                   mode,
+        int                   n_ctx) {
     GGML_ASSERT(n_past >= 0);
     GGML_ASSERT((mode & 4) == 0 && "ggml_rope_back() for ChatGLM not implemented yet");
 
@@ -7043,12 +7044,13 @@ struct ggml_tensor * ggml_rope_back(
 
     ggml_scratch_save(ctx);
 
-    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3);
+    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 4);
     ggml_set_name(b, "n_past, n_dims, mode");
 
     ((int32_t *) b->data)[0] = n_past;
     ((int32_t *) b->data)[1] = n_dims;
     ((int32_t *) b->data)[2] = mode;
+    ((int32_t *) b->data)[3] = n_ctx;
 
     ggml_scratch_load(ctx);
 
@@ -15740,13 +15742,15 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     const int n_past = ((int32_t *) src1->data)[0];
                     const int n_dims = ((int32_t *) src1->data)[1];
                     const int mode   = ((int32_t *) src1->data)[2];
+                    const int n_ctx  = ((int32_t *) src1->data)[3];
                     src0->grad = ggml_add_impl(ctx,
                             src0->grad,
                             ggml_rope_back(ctx,
                                 tensor->grad,
                                 n_past,
                                 n_dims,
-                                mode),
+                                mode,
+                                n_ctx),
                             inplace);
                 }
                 if (src1->grad) {
@@ -15757,7 +15761,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 if (src0->grad) {
                     assert(src1->type == GGML_TYPE_I32);
-                    assert(ggml_nelements(src1) == 3);
+                    assert(ggml_nelements(src1) == 4);
                     const int n_past = ((int32_t *) src1->data)[0];
                     const int n_dims = ((int32_t *) src1->data)[1];
                     const int mode   = ((int32_t *) src1->data)[2];
diff --git a/ggml.h b/ggml.h
index 24856a255..5023b1652 100644
--- a/ggml.h
+++ b/ggml.h
@@ -1128,9 +1128,9 @@ extern "C" {
             int                   n_past,
             int                   n_dims,
             int                   mode,
+            int                   n_ctx,
             float                 freq_base,
-            float                 freq_scale,
-            int                   n_ctx);
+            float                 freq_scale);
 
     // rotary position embedding backward, i.e compute dx from dy
     // a - dy
@@ -1139,7 +1139,8 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_past,
             int                   n_dims,
-            int                   mode);
+            int                   mode,
+            int                   n_ctx);
 
     // alibi position embedding
     // in-place, returns view(a)
diff --git a/llama.cpp b/llama.cpp
index 3b0024e12..0a381afd5 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -1452,11 +1452,11 @@ static bool llama_eval_internal(
             offload_func_kq(tmpq);
             ggml_set_name(tmpq, "tmpq");
 
-            struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0, freq_base, freq_scale, 0);
+            struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0, 0, freq_base, freq_scale);
             offload_func_kq(Kcur);
             ggml_set_name(Kcur, "Kcur");
 
-            struct ggml_tensor * Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0, freq_base, freq_scale, 0);
+            struct ggml_tensor * Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0, 0, freq_base, freq_scale);
             offload_func_kq(Qcur);
             ggml_set_name(Qcur, "Qcur");
 

From 03e566977b277937c5f706180171c5d12b597b0b Mon Sep 17 00:00:00 2001
From: Ikko Eltociear Ashimine <eltociear@gmail.com>
Date: Fri, 21 Jul 2023 20:53:07 +0900
Subject: [PATCH 16/20] examples :  fix typo in minigpt4.py (#2298)

promt -> prompt
---
 examples/embd-input/minigpt4.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/examples/embd-input/minigpt4.py b/examples/embd-input/minigpt4.py
index 8e98f8517..15c9b77c0 100644
--- a/examples/embd-input/minigpt4.py
+++ b/examples/embd-input/minigpt4.py
@@ -64,7 +64,7 @@ class MiniGPT4(Blip2Base):
         self.max_txt_len = max_txt_len
         self.end_sym = end_sym
         self.model = MyModel(["main", *args])
-        # system promt
+        # system prompt
         self.model.eval_string("Give the following image: <Img>ImageContent</Img>. "
            "You will be able to see the image once I provide it to you. Please answer my questions."
            "###")

From 0db14fef06836caaa13cc123c0a24dc598bdb9f0 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 21 Jul 2023 15:16:55 +0300
Subject: [PATCH 17/20] ggml : fix the rope fix
 (513f8619535a64fa9ace808cdcbcf66211535f5c)

---
 ggml.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml.c b/ggml.c
index 7ecabc5de..6055da867 100644
--- a/ggml.c
+++ b/ggml.c
@@ -12379,7 +12379,7 @@ static void ggml_compute_forward_rope_back_f32(
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
     assert(src1->type == GGML_TYPE_I32);
-    assert(ggml_nelements(src1) == 3);
+    assert(ggml_nelements(src1) == 4);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;

From 4d76a5f49b9b5382dba5d13d92edb9159536c225 Mon Sep 17 00:00:00 2001
From: Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Date: Fri, 21 Jul 2023 17:05:30 +0300
Subject: [PATCH 18/20] Faster Q3_K implementation on Metal (#2307)

* Faster Q3_K on Metal

* Additional Q3_K speedup on Metal

* Q3_K for QK_K = 64

* Better Q3_K for QK_K = 64

21.6 ms/t -> 21.1 ms/t

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
---
 ggml-metal.m     |  15 ++--
 ggml-metal.metal | 192 ++++++++++++++++++++++++++++-------------------
 2 files changed, 125 insertions(+), 82 deletions(-)

diff --git a/ggml-metal.m b/ggml-metal.m
index 135bda9fc..2810fa2a8 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -685,8 +685,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
-                                            nth1 = 16;
+                                            nth0 = 2;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
                                         } break;
                                     case GGML_TYPE_Q4_K:
@@ -743,15 +743,18 @@ void ggml_metal_graph_compute(
                                     src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
+                                else if (src0t == GGML_TYPE_Q3_K) {
+#ifdef GGML_QKK_64
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#else
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#endif
+                                }
                                 else if (src0t == GGML_TYPE_Q5_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q6_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                }
-                                else if (src0t == GGML_TYPE_Q3_K) {
-                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
                                     [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
diff --git a/ggml-metal.metal b/ggml-metal.metal
index 97f5c10ba..5a9a6d842 100644
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
@@ -351,7 +351,7 @@ kernel void kernel_rms_norm(
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
     // broadcast, simd group number is ntg / 32
-    for (int i = ntg / 32 / 2; i > 0; i /= 2) {
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
        if (tpitg < i) {
            sum[tpitg] += sum[tpitg + i];
        }
@@ -1339,6 +1339,7 @@ kernel void kernel_mul_mat_q2_K_f32(
     }
 }
 
+#if QK_K == 256
 kernel void kernel_mul_mat_q3_K_f32(
         device const  void * src0,
         device const float * src1,
@@ -1347,40 +1348,41 @@ kernel void kernel_mul_mat_q3_K_f32(
         constant   int64_t & ne10,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
-        threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2 tpitg[[thread_position_in_threadgroup]],
-        uint2  tptg[[threads_per_threadgroup]]) {
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const int nb = ne00/QK_K;
 
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
 
-    device const block_q3_K * x = (device const block_q3_K *) src0 + r0*nb;
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
+
+    device const block_q3_K * x = (device const block_q3_K *) src0 + first_row*nb;
     device const float     * yy = (device const float      *) src1 + r1*ne10;
 
-    const int nth = tptg.x*tptg.y;
-    const int ith = tptg.y*tpitg.x + tpitg.y;
-
-#if QK_K == 256
-
-    const uint8_t m3 = 3;
-    const int8_t  m4 = 4;
+    float yl[16];
 
     const uint16_t kmask1 = 0x0303;
     const uint16_t kmask2 = 0x0f0f;
 
-    const int tid = tpitg.y;        // expecting 16
+    const int tid = tiisg/2;
+    const int ix  = tiisg%2;
     const int ip  = tid/8;          // 0 or 1
     const int il  = tid/2 - 4*ip;   // 0...3
     const int ir  = tid%2;
     const int n   = 8;
     const int l0  = n*ir;
 
-    const uint8_t m = 1 << (4*ip + il);
+    const uint16_t m1 = 1 << (4*ip + il);
+    const uint16_t m2 = m1 << 8;
 
     const int shift = 2*il;
+    const uint16_t qm1 = 0x0003 << shift;
+    const uint16_t qm2 = 0x0300 << shift;
+    const int32_t v1 = 4 << shift;
+    const int32_t v2 = 1024 << shift;
 
     const uint16_t s_shift1 = 4*ip;
     const uint16_t s_shift2 = s_shift1 + 2*(il/2);
@@ -1389,93 +1391,132 @@ kernel void kernel_mul_mat_q3_K_f32(
     const int q_offset = 32*ip + l0;
     const int y_offset = 128*ip + 32*il + l0;
 
-    //float sumf = 0;
-    float sumf1 = 0, sumf2 = 0;
-    for (int i = tpitg.x; i < nb; i += tptg.x) {
+    const int step = sizeof(block_q3_K) * nb / 2;
 
-        const float d_all = (float)(x[i].d);
+    device const float * y1 = yy + ix*QK_K + y_offset;
 
-        device const uint8_t * q = x[i].qs + q_offset;
-        device const uint8_t * h = x[i].hmask + l0;
-        device const float   * y = yy + i * QK_K + y_offset;
+    float sumf1[2] = {0.f}, sumf2[2] = {0.f};
+    for (int i = ix; i < nb; i += 2) {
 
-        device const uint16_t * a = (device const uint16_t *)x[i].scales;
-        const char2 scales = as_type<char2>((uint16_t)(((a[il] >> s_shift1) & kmask2) | (((a[ik] >> s_shift2) & kmask1) << 4)));
-
-        float s = 0;
-        for (int l = 0; l < n; ++l) {
-            s += y[l+ 0] * ((int8_t)((q[l+ 0] >> shift) & m3) - ((h[l+ 0] & m) ? 0 : m4));
+        for (int l = 0; l < 8; ++l) {
+            yl[l+0] = y1[l+ 0];
+            yl[l+8] = y1[l+16];
         }
-        float d = d_all * s;
-        sumf1 += d * scales[0];
-        sumf2 += d;
-        //sumf += d_all * s * (scales[0] - 32);
 
-        s = 0;
-        for (int l = 0; l < n; ++l) {
-            s += y[l+16] * ((int8_t)((q[l+16] >> shift) & m3) - ((h[l+16] & m) ? 0 : m4));
+        device const uint16_t * q = (device const uint16_t *)(x[i].qs + q_offset);
+        device const uint16_t * h = (device const uint16_t *)(x[i].hmask + l0);
+        device const uint16_t * a = (device const uint16_t *)(x[i].scales);
+        device const half * dh = &x[i].d;
+
+        for (int row = 0; row < 2; ++row) {
+
+            const float d_all = (float)dh[0];
+            const char2 scales = as_type<char2>((uint16_t)(((a[il] >> s_shift1) & kmask2) | (((a[ik] >> s_shift2) & kmask1) << 4)));
+
+            float s1 = 0, s2 = 0;
+            for (int l = 0; l < n; l += 2) {
+                const uint16_t qs = q[l/2];
+                s1 += yl[l+0] * ((int32_t)(qs & qm1) - ((h[l/2] & m1) ? 0 : v1));
+                s2 += yl[l+1] * ((int32_t)(qs & qm2) - ((h[l/2] & m2) ? 0 : v2));
+            }
+            float d = d_all * (s1 + 1.f/256.f * s2);
+            sumf1[row] += d * scales[0];
+            sumf2[row] += d;
+
+            s1 = s2 = 0;
+            for (int l = 0; l < n; l += 2) {
+                const uint16_t qs = q[l/2+8];
+                s1 += yl[l+8] * ((int32_t)(qs & qm1) - ((h[l/2+8] & m1) ? 0 : v1));
+                s2 += yl[l+9] * ((int32_t)(qs & qm2) - ((h[l/2+8] & m2) ? 0 : v2));
+            }
+            d = d_all * (s1 + 1.f/256.f * s2);
+            sumf1[row] += d * scales[1];
+            sumf2[row] += d;
+
+            q  += step;
+            h  += step;
+            a  += step;
+            dh += step;
+
         }
-        d = d_all * s;
-        sumf1 += d * scales[1];
-        sumf2 += d;
-        //sumf += d_all * s * (scales[1] - 32);
+
+        y1 += 2 * QK_K;
 
     }
 
-    //sum[ith] = sumf;
-    sum[ith] = sumf1 - 32.f*sumf2;
+    for (int row = 0; row < 2; ++row) {
+        const float sumf = (sumf1[row] - 32.f*sumf2[row]) / (1 << shift);
+        const float tot = simd_sum(sumf);
+        if (tiisg == 0) {
+            dst[r1*ne0 + first_row + row] = tot;
+        }
+    }
+}
 #else
-    const int il = 4 * tpitg.x;  // 0, 4, 8, 12
+kernel void kernel_mul_mat_q3_K_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne10,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+
+    const int row = 2 * r0 + sgitg;
+
+    device const block_q3_K * x = (device const block_q3_K *) src0 + row*nb;
+    device const float     * yy = (device const float      *) src1 + r1*ne10;
+    const int ix = tiisg/4;
+    const int il = 4 * (tiisg%4);// 0, 4, 8, 12
     const int im = il/8;         // 0, 0, 1, 1
     const int in = il%8;         // 0, 4, 0, 4
 
-    float sumf = 0;
+    float2 sum = {0.f, 0.f};
 
-    for (int i = tpitg.y; i < nb; i += tptg.y) {
+    for (int i = ix; i < nb; i += 8) {
 
         const float d_all = (float)(x[i].d);
 
-        device const uint8_t * q = x[i].qs + il;
-        device const uint8_t * h = x[i].hmask + in;
-        device const float   * y = yy + i * QK_K + il;
+        device const uint16_t * q = (device const uint16_t *)(x[i].qs + il);
+        device const uint16_t * h = (device const uint16_t *)(x[i].hmask + in);
+        device const uint16_t * s = (device const uint16_t *)(x[i].scales);
+        device const float    * y = yy + i * QK_K + il;
 
-        const float d1 = d_all * ((x[i].scales[0] & 0xF) - 8);
-        const float d2 = d_all * ((x[i].scales[0] >>  4) - 8);
-        const float d3 = d_all * ((x[i].scales[1] & 0xF) - 8);
-        const float d4 = d_all * ((x[i].scales[1] >>  4) - 8);
+        const float d1 = d_all * ((int32_t)(s[0] & 0x000F) - 8);
+        const float d2 = d_all * ((int32_t)(s[0] & 0x00F0) - 128) * 1.f/64.f;
+        const float d3 = d_all * ((int32_t)(s[0] & 0x0F00) - 2048) * 1.f/4096.f;
+        const float d4 = d_all * ((int32_t)(s[0] & 0xF000) - 32768) * 1.f/262144.f;
 
-        for (int l = 0; l < 4; ++l) {
-            const uint8_t hm = h[l] >> im;
-            sumf += y[l+ 0] * d1 * ((int8_t)((q[l+0] >> 0) & 3) - ((hm & 0x01) ? 0 : 4))
-                  + y[l+16] * d2 * ((int8_t)((q[l+0] >> 2) & 3) - ((hm & 0x04) ? 0 : 4))
-                  + y[l+32] * d3 * ((int8_t)((q[l+0] >> 4) & 3) - ((hm & 0x10) ? 0 : 4))
-                  + y[l+48] * d4 * ((int8_t)((q[l+0] >> 6) & 3) - ((hm & 0x40) ? 0 : 4));
+        for (int l = 0; l < 4; l += 2) {
+            const uint16_t hm = h[l/2] >> im;
+            sum[0] += y[l+ 0] * d1 * ((int32_t)(q[l/2] & 0x0003) - ((hm & 0x0001) ? 0 :  4))
+                    + y[l+16] * d2 * ((int32_t)(q[l/2] & 0x000c) - ((hm & 0x0004) ? 0 : 16))
+                    + y[l+32] * d3 * ((int32_t)(q[l/2] & 0x0030) - ((hm & 0x0010) ? 0 : 64))
+                    + y[l+48] * d4 * ((int32_t)(q[l/2] & 0x00c0) - ((hm & 0x0040) ? 0 : 256));
+            sum[1] += y[l+ 1] * d1 * ((int32_t)(q[l/2] & 0x0300) - ((hm & 0x0100) ? 0 : 1024))
+                    + y[l+17] * d2 * ((int32_t)(q[l/2] & 0x0c00) - ((hm & 0x0400) ? 0 : 4096))
+                    + y[l+33] * d3 * ((int32_t)(q[l/2] & 0x3000) - ((hm & 0x1000) ? 0 : 16384))
+                    + y[l+49] * d4 * ((int32_t)(q[l/2] & 0xc000) - ((hm & 0x4000) ? 0 : 65536));
         }
 
     }
+    const float sumf = sum[0] + sum[1] * 1.f/256.f;
 
-    sum[ith] = sumf;
-
-#endif
-
-    //
-    // Accumulate the sum from all threads in the threadgroup
-    //
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%4 == 0) {
-        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith%16 == 0) {
-        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (ith == 0) {
-        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
-        dst[r1*ne0 + r0] = sum[0];
+    const float tot = simd_sum(sumf);
+    if (tiisg == 0) {
+        dst[r1*ne0 + row] = tot;
     }
 
 }
+#endif
 
 #if QK_K == 256
 kernel void kernel_mul_mat_q4_K_f32(
@@ -1773,7 +1814,6 @@ kernel void kernel_mul_mat_q5_K_f32(
 
     for (int i = ix; i < nb; i += 8) {
 
-        float4 sumy = {0.f, 0.f, 0.f, 0.f};
         for (int l = 0; l < 4; ++l) {
             yl[l+0] = y[l+ 0];
             yl[l+4] = y[l+16];

From d924522a46c5ef097af4a88087d91673e8e87e4d Mon Sep 17 00:00:00 2001
From: Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Date: Fri, 21 Jul 2023 17:27:51 +0300
Subject: [PATCH 19/20] Custom RoPE + bettter memory management for CUDA
 (#2295)

* Custom RoPE + bettter memory management for CUDA

* Adjusted look ahead in ggml_cuda_pool_malloc to 5%

This is sufficient it seems.
We end up using about 200 MB less VRAM that way when running
the 13B model with context 8192.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
---
 ggml-cuda.cu | 60 ++++++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 49 insertions(+), 11 deletions(-)

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 6537897b9..c07b54611 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -2423,20 +2423,53 @@ static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
     scoped_spin_lock lock(g_cuda_pool_lock);
     int id;
     CUDA_CHECK(cudaGetDevice(&id));
-
+#ifdef DEBUG_CUDA_MALLOC
+    int nnz = 0;
+    size_t max_size = 0, tot_size = 0;
+#endif
+    size_t best_diff = 1ull << 36;
+    int ibest = -1;
     for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
         cuda_buffer& b = g_cuda_buffer_pool[id][i];
-        if (b.size >= size && b.ptr != nullptr) {
-            void * ptr = b.ptr;
-            *actual_size = b.size;
-            b.ptr = nullptr;
-            b.size = 0;
-            return ptr;
+        if (b.ptr != nullptr) {
+#ifdef DEBUG_CUDA_MALLOC
+            ++nnz;
+            tot_size += b.size;
+            if (b.size > max_size) max_size = b.size;
+#endif
+            if (b.size >= size) {
+                size_t diff = b.size - size;
+                if (diff < best_diff) {
+                    best_diff = diff;
+                    ibest = i;
+                    if (!best_diff) {
+                        void * ptr = b.ptr;
+                        *actual_size = b.size;
+                        b.ptr = nullptr;
+                        b.size = 0;
+                        return ptr;
+                    }
+                }
+            }
         }
     }
+    if (ibest >= 0) {
+        cuda_buffer& b = g_cuda_buffer_pool[id][ibest];
+        void * ptr = b.ptr;
+        *actual_size = b.size;
+        b.ptr = nullptr;
+        b.size = 0;
+        return ptr;
+    }
+#ifdef DEBUG_CUDA_MALLOC
+    fprintf(stderr, "%s: %d buffers, max_size = %u MB, tot_size = %u MB, requested %u MB\n", __func__, nnz,
+            (uint32_t)(max_size/1024/1024), (uint32_t)(tot_size/1024/1024), (uint32_t)(size/1024/1024));
+#endif
     void * ptr;
-    CUDA_CHECK(cudaMalloc((void **) &ptr, size));
-    *actual_size = size;
+    size_t look_ahead_size = (size_t) (1.05 * size);
+    look_ahead_size = 256 * ((look_ahead_size + 255)/256);
+    CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size));
+    *actual_size = look_ahead_size;
     return ptr;
 }
 
@@ -2955,8 +2988,13 @@ inline void ggml_cuda_op_rope(
     const int mode   = ((int32_t *) src1->data)[2];
     const int n_ctx  = ((int32_t *) src1->data)[3];
 
-    const float theta_scale = powf(10000.0, -2.0f/n_dims);
-    const float p = ((mode & 1) == 0 ? n_past + i02 : i02);
+    // RoPE alteration for extended context
+    float freq_base, freq_scale;
+    memcpy(&freq_base,  (int32_t *) src1->data + 4, sizeof(float));
+    memcpy(&freq_scale, (int32_t *) src1->data + 5, sizeof(float));
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float p = (((mode & 1) == 0 ? n_past + i02 : i02)) * freq_scale;
 
     bool is_glm = mode & 4;
 

From 7d5f18468ceabd7a38f414f9f21b26b0c137f994 Mon Sep 17 00:00:00 2001
From: Richard Roberson <richardr1126@gmail.com>
Date: Fri, 21 Jul 2023 13:01:10 -0600
Subject: [PATCH 20/20] examples : add easy python script to create quantized
 (k-bit support) GGML models from local HF Transformer models (#2311)

* Resync my fork with new llama.cpp commits

* examples : rename to use dash instead of underscore

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---
 examples/make-ggml.py | 92 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 92 insertions(+)
 create mode 100644 examples/make-ggml.py

diff --git a/examples/make-ggml.py b/examples/make-ggml.py
new file mode 100644
index 000000000..f63d9fc22
--- /dev/null
+++ b/examples/make-ggml.py
@@ -0,0 +1,92 @@
+"""
+This script converts Hugging Face llama models to GGML and quantizes them.
+
+Usage:
+python make-ggml.py --model {model_dir_or_hf_repo_name} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
+
+Arguments:
+- --model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- --outname: (Optional) The name of the output model. If not specified, the last part of the model directory path or the Hugging Face model repo name will be used.
+- --outdir: (Optional) The directory where the output model(s) will be stored. If not specified, '../models/{outname}' will be used.
+- --quants: (Optional) The types of quantization to apply. This should be a space-separated list. The default is 'Q4_K_M Q5_K_S'.
+- --keep_fp16: (Optional) If specified, the FP16 model will not be deleted after the quantized models are created.
+
+Quant types:
+- Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M
+- Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L
+- Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M
+- Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M
+- Q2_K: smallest, extreme quality loss - not recommended
+- Q3_K: alias for Q3_K_M
+- Q3_K_S: very small, very high quality loss
+- Q3_K_M: very small, very high quality loss
+- Q3_K_L: small, substantial quality loss
+- Q4_K: alias for Q4_K_M
+- Q4_K_S: small, significant quality loss
+- Q4_K_M: medium, balanced quality - recommended
+- Q5_K: alias for Q5_K_M
+- Q5_K_S: large, low quality loss - recommended
+- Q5_K_M: large, very low quality loss - recommended
+- Q6_K: very large, extremely low quality loss
+- Q8_0: very large, extremely low quality loss - not recommended
+- F16: extremely large, virtually no quality loss - not recommended
+- F32: absolutely huge, lossless - not recommended
+"""
+import subprocess
+subprocess.run(f"pip install huggingface-hub==0.16.4", shell=True, check=True)
+
+import argparse
+import os
+from huggingface_hub import snapshot_download
+
+def main(model, outname, outdir, quants, keep_fp16):
+    ggml_version = "v3"
+
+    if not os.path.isdir(model):
+        print(f"Model not found at {model}. Downloading...")
+        try:
+            if outname is None:
+                outname = model.split('/')[-1]
+            model = snapshot_download(repo_id=model, cache_dir='../models/hf_cache')
+        except Exception as e:
+            raise Exception(f"Could not download the model: {e}")
+
+    if outdir is None:
+        outdir = f'../models/{outname}'
+
+    if not os.path.isfile(f"{model}/config.json"):
+        raise Exception(f"Could not find config.json in {model}")
+
+    os.makedirs(outdir, exist_ok=True)
+
+    print("Building llama.cpp")
+    subprocess.run(f"cd .. && make quantize", shell=True, check=True)
+
+    fp16 = f"{outdir}/{outname}.ggml{ggml_version}.fp16.bin"
+
+    print(f"Making unquantised GGML at {fp16}")
+    if not os.path.isfile(fp16):
+        subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
+    else:
+        print(f"Unquantised GGML already exists at: {fp16}")
+
+    print("Making quants")
+    for type in quants:
+        outfile = f"{outdir}/{outname}.ggml{ggml_version}.{type}.bin"
+        print(f"Making {type} : {outfile}")
+        subprocess.run(f"../quantize {fp16} {outfile} {type}", shell=True, check=True)
+
+    if not keep_fp16:
+        os.remove(fp16)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Convert/Quantize HF to GGML. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
+    parser.add_argument('--model', required=True, help='Downloaded model dir or Hugging Face model repo name')
+    parser.add_argument('--outname', default=None, help='Output model(s) name')
+    parser.add_argument('--outdir', default=None, help='Output directory')
+    parser.add_argument('--quants', nargs='*', default=["Q4_K_M", "Q5_K_S"], help='Quant types')
+    parser.add_argument('--keep_fp16', action='store_true', help='Keep fp16 model', default=False)
+
+    args = parser.parse_args()
+
+    main(args.model, args.outname, args.outdir, args.quants, args.keep_fp16)