cmake : link threads publicly to ggml (#1042)

* fix: ld link test-tokenizer-0 error

```
cmake3 --build . --config Release
[  5%] Built target ggml
[ 16%] Built target llama
[ 22%] Linking CXX executable ../bin/test-tokenizer-0
../libllama.a(ggml.c.o)：在函数‘ggml_graph_compute’中：
ggml.c:(.text+0xf2db)：对‘pthread_create’未定义的引用
ggml.c:(.text+0xf9d4)：对‘pthread_join’未定义的引用
collect2: error: ld returned 1 exit status
gmake[2]: *** [bin/test-tokenizer-0] 错误 1
gmake[1]: *** [tests/CMakeFiles/test-tokenizer-0.dir/all] 错误 2
gmake: *** [all] 错误 2
```

* Update CMakeLists.txt

* Update CMakeLists.txt

* Update CMakeLists.txt

CMakeLists.txt

@@ -307,7 +307,7 @@ add_library(ggml OBJECT
 
 target_include_directories(ggml PUBLIC .)
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
-target_link_libraries(ggml PRIVATE Threads::Threads ${LLAMA_EXTRA_LIBS})
+target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
 if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()

SHA256SUMS

@@ -1,12 +1,27 @@
 700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d  models/7B/consolidated.00.pth
+666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847  models/7B/ggml-model-f16.bin
+fcb7664c2e69776920b526362a243e912f73c36b1ec892eb354bab940f5edb5a  models/7B/ggml-model-q4_0.bin
+cc061458339a3eb8bcecbf0a825e9924fb7d1a8150f63cd5d091caa99215aafe  models/7B/ggml-model-q4_1.bin
+1bc7484c24a87612726d756f1761890e7acf5f412e23378577ce50fbe789b5b8  models/7B/ggml-model-q4_2.bin
+3429bf198ec771886cf81a574df45245f3ebf04f0ce0956b73ef5d0ab01ff48b  models/7B/ggml-model-q4_3.bin
 7e89e242ddc0dd6f060b43ca219ce8b3e8f08959a72cb3c0855df8bb04d46265  models/7B/params.json
 745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08  models/13B/consolidated.00.pth
 d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085  models/13B/consolidated.01.pth
+2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808  models/13B/ggml-model-f16.bin
+4b69e4d6b6e3275230955997b90407fceca7e5ab3daf2e63a2c9e7270a8e1e3e  models/13B/ggml-model-q4_0.bin
+d9581b5b88e5622532fe897c9f9b0e67a317d22dd27a6f90fa4ab8c6d23ccdbb  models/13B/ggml-model-q4_1.bin
+8d55a2077317ec9a928c7851d6a43e08e51f7e9e08360f2a7a7e1deefea3134f  models/13B/ggml-model-q4_2.bin
+4208cdec9788ffa48dc1a17af2c36a0299f5bf3eb0e2b87889dda7fad591fca3  models/13B/ggml-model-q4_3.bin
 4ab77bec4d4405ccb66a97b282574c89a94417e3c32e5f68f37e2876fc21322f  models/13B/params.json
 e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067  models/30B/consolidated.00.pth
 4e077b7136c7ae2302e954860cf64930458d3076fcde9443f4d0e939e95903ff  models/30B/consolidated.01.pth
 24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378  models/30B/consolidated.02.pth
 1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b  models/30B/consolidated.03.pth
+7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37  models/30B/ggml-model-f16.bin
+7a679908ce31c9d6ae2e38d6059bcd4d0ad3a870cd58cc1c8f7b36f2b2f51c73  models/30B/ggml-model-q4_0.bin
+7b75ac615fa369ee593493a7e6ef87542bf0350255db928b22c5a24f6d598bcd  models/30B/ggml-model-q4_1.bin
+2c82b4954a94a6a284f452f6011c1e4f0d20362c194a0b1eb5737f5fd8a20fb3  models/30B/ggml-model-q4_2.bin
+a6188660199dbcb8d5658abe7d89169869e50423494385830d9e6b330ea7fc33  models/30B/ggml-model-q4_3.bin
 2c07118ea98d69dbe7810d88520e30288fa994751b337f8fca02b171955f44cb  models/30B/params.json
 135c563f6b3938114458183afb01adc9a63bef3d8ff7cccc3977e5d3664ecafe  models/65B/consolidated.00.pth
 9a600b37b19d38c7e43809485f70d17d1dc12206c07efa83bc72bb498a568bde  models/65B/consolidated.01.pth
@@ -16,5 +31,10 @@ e7babf7c5606f165a3756f527cb0fedc4f83e67ef1290391e52fb1cce5f26770  models/65B/con
 a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78  models/65B/consolidated.05.pth
 72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b  models/65B/consolidated.06.pth
 d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638  models/65B/consolidated.07.pth
+60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0  models/65B/ggml-model-f16.bin
+c671fe1bce71499ac732ec999770ebe53ac486623a7891e42c9dfdb6962d2c64  models/65B/ggml-model-q4_0.bin
+4743a28aac3e5f32a6e838a815f51d3779de44fbbe251d745251e66c23c5950f  models/65B/ggml-model-q4_1.bin
+4a145a210c56982389b1ed34387e0590c3e0d7325fa9be4f2284fe4d244a3633  models/65B/ggml-model-q4_2.bin
+305e91a4608b4f627b9b8ad5b4af75187d2684254bfd76dcb9db571618ef293c  models/65B/ggml-model-q4_3.bin
 999ed1659b469ccc2a941714c0a9656fa571d17c9f7c8c7589817ca90edef51b  models/65B/params.json
 9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347  models/tokenizer.model

examples/main/main.cpp

@@ -264,7 +264,7 @@ int main(int argc, char ** argv) {
             // infinite text generation via context swapping
             // if we run out of context:
             // - take the n_keep first tokens from the original prompt (via n_past)
-            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in a batch
+            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
             if (n_past + (int) embd.size() > n_ctx) {
                 const int n_left = n_past - params.n_keep;
 
@@ -282,13 +282,21 @@ int main(int argc, char ** argv) {
                 //printf("\n---\n");
             }
 
-            if (llama_eval(ctx, embd.data(), embd.size(), n_past, params.n_threads)) {
-                fprintf(stderr, "%s : failed to eval\n", __func__);
-                return 1;
+            // evaluate tokens in batches
+            // embd is typically prepared beforehand to fit within a batch, but not always
+            for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
+                int n_eval = (int) embd.size() - i;
+                if (n_eval > params.n_batch) {
+                    n_eval = params.n_batch;
+                }
+                if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) {
+                    fprintf(stderr, "%s : failed to eval\n", __func__);
+                    return 1;
+                }
+                n_past += n_eval;
             }
         }
 
-        n_past += embd.size();
         embd.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {

examples/perplexity/perplexity.cpp

@@ -53,7 +53,13 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
         auto end_t = std::chrono::high_resolution_clock::now();
         if (i == 0) {
             const float seconds = std::chrono::duration<float>(end_t - start_t).count();
-            printf("%.2f seconds per pass - ETA %.2f hours\n", seconds, (seconds * seq_count) / (60.0*60.0));
+            printf("%.2f seconds per pass - ETA ", seconds);
+            int total_seconds = (int)(seconds * seq_count);
+            if (total_seconds >= 60*60) {
+                printf("%d hours ", total_seconds / (60*60));
+                total_seconds = total_seconds % (60*60);
+            }
+            printf("%d minutes\n", total_seconds / 60);
         }
         // We get the logits for all the tokens in the context window (params.n_ctx)
         // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,

ggml.c

@@ -657,9 +657,10 @@ static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong
 #define QK8_0 32
 typedef struct {
     float   d;          // delta
+    float   s;          // d * sum(qs[i])
     int8_t  qs[QK8_0];  // quants
 } block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+static_assert(sizeof(block_q8_0) == 2*sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
 
 
 // reference implementation for deterministic creation of model files
@@ -1299,13 +1300,39 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r
 
         y[i].d = d;
 
+        int sum = 0;
         for (int l = 0; l < QK8_0; ++l) {
             const float v = x[i*QK8_0 + l]*id;
             y[i].qs[l] = roundf(v);
+            sum += y[i].qs[l];
         }
+        y[i].s = d * sum;
     }
 }
 
+#ifdef __AVX2__
+// There is no better way of doing this?
+// I guess not, AVX is not very good at horizontal sums.
+// The commented solution for a hotrizontal sum was suggested by @pubby as being slightly
+// faster than the solution below. As I don't have an AVX2 system handt right now to test,
+// keeping the original.
+// TODO: Please try and if it does make a differece, uncomment and remove the implementation below.
+//static inline float horizontal_sum(__m256i a) {
+//    __m256i b = _mm256_castps_si256(_mm256_movehdup_ps(_mm256_castsi256_ps(a)));
+//    __m256i sum = _mm256_add_epi32(a, b);
+//    __m256i hi = _mm256_unpackhi_epi64(sum, sum);
+//    sum = _mm256_add_epi32(sum, hi);
+//    return _mm256_cvtsi256_si32(sum) + _mm256_extract_epi32(sum, 4);
+//}
+static inline float horizontal_sum(__m256i a) {
+    __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extracti128_si256(a, 1));
+    __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
+    __m128i sum64 = _mm_add_epi32(hi64, sum128);
+    __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+#endif
+
 static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) {
     assert(k % QK8_0 == 0);
     const int nb = k / QK8_0;
@@ -1332,6 +1359,8 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
 
         y[i].d = d;
 
+        int32x4_t accv = vdupq_n_s32(0);
+
         for (int l = 0; l < 8; l++) {
             const float32x4_t v  = vmulq_n_f32(srcv[l], id);
             const int32x4_t   vi = vcvtnq_s32_f32(v);
@@ -1340,7 +1369,11 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
             y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1);
             y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2);
             y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3);
+
+            accv = vaddq_s32(accv, vi);
         }
+        int32_t sum = vaddvq_s32(accv);
+        y[i].s = d * sum;
     }
 #elif defined(__AVX2__) || defined(__AVX__)
     for (int i = 0; i < nb; i++) {
@@ -1388,6 +1421,10 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
         __m256i i3 = _mm256_cvtps_epi32( v3 );
 
 #if defined(__AVX2__)
+
+        // Compute the sum of the quants and set y[i].s
+        y[i].s = d * horizontal_sum(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
+
         // Convert int32 to int16
         i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
         i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
@@ -1430,6 +1467,14 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
     // scalar
     quantize_row_q8_0_reference(x, y, k);
 #endif
+#if defined __AVX__
+    // TODO: vectorize this
+    for (int i=0; i<nb; ++i) {
+        int sum = 0;
+        for (int l=0; l<QK8_0; ++l) sum += y[i].qs[l];
+        y[i].s = y[i].d * sum;
+    }
+#endif
 }
 
 static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, int k) {
@@ -2372,14 +2417,17 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
+    float sum8 = 0;
+
     for (int i = 0; i < nb; i += 2) {
         const block_q4_0 * restrict x0 = &x[i + 0];
         const block_q4_0 * restrict x1 = &x[i + 1];
         const block_q8_0 * restrict y0 = &y[i + 0];
         const block_q8_0 * restrict y1 = &y[i + 1];
 
+        sum8 += x0->d * y0->s + x1->d * y1->s;
+
         const uint8x16_t m4b   = vdupq_n_u8(0xf);
-        const int8x16_t  s8b   = vdupq_n_s8(0x8);
 
         const uint8x16_t v0_0 = vld1q_u8(x0->qs);
         const uint8x16_t v0_1 = vld1q_u8(x1->qs);
@@ -2390,12 +2438,6 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
         const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
 
-        // sub 8
-        const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b);
-        const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b);
-        const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
-        const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
-
         // load y
         const int8x16_t v1_0l = vld1q_s8(y0->qs);
         const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
@@ -2410,21 +2452,21 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
 
 #if defined(__ARM_FEATURE_DOTPROD)
         // dot product into int32x4_t
-        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls), v0_0hs, v1_0hs);
-        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls), v0_1hs, v1_1hs);
+        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0ls), v0_0h, v1_0hs);
+        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1ls), v0_1h, v1_1hs);
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d);
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d);
 #else
-        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls));
-        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls));
-        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0hs));
-        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0hs));
+        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0ls));
+        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0ls));
+        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0hs));
+        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0hs));
 
-        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1ls));
-        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1ls));
-        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs));
-        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1hs));
+        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1ls));
+        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1ls));
+        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1hs));
+        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1hs));
 
         const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
         const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
@@ -2436,7 +2478,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
 #endif
     }
 
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) - 8 * sum8;
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -2569,12 +2611,16 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
+    float summs = 0;
+
     for (int i = 0; i < nb; i += 2) {
         const block_q4_1 * restrict x0 = &x[i + 0];
         const block_q4_1 * restrict x1 = &x[i + 1];
         const block_q8_0 * restrict y0 = &y[i + 0];
         const block_q8_0 * restrict y1 = &y[i + 1];
 
+        summs += x0->m * y0->s + x1->m * y1->s;
+
         const uint8x16_t m4b = vdupq_n_u8(0xf);
 
         const uint8x16_t v0_0 = vld1q_u8(x0->qs);
@@ -2598,17 +2644,6 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
         const int8x16_t v1_1ls = vuzp1q_s8(v1_1l, v1_1h);
         const int8x16_t v1_1hs = vuzp2q_s8(v1_1l, v1_1h);
 
-        const int16x8_t s0i = vaddq_s16(
-                        vaddq_s16(vmovl_s8(vget_low_s8(v1_0ls)), vmovl_s8(vget_high_s8(v1_0ls))),
-                        vaddq_s16(vmovl_s8(vget_low_s8(v1_0hs)), vmovl_s8(vget_high_s8(v1_0hs))));
-
-        const int16x8_t s1i = vaddq_s16(
-                        vaddq_s16(vmovl_s8(vget_low_s8(v1_1ls)), vmovl_s8(vget_high_s8(v1_1ls))),
-                        vaddq_s16(vmovl_s8(vget_low_s8(v1_1hs)), vmovl_s8(vget_high_s8(v1_1hs))));
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(s0i), vget_high_s16(s0i))), x0->m*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(s1i), vget_high_s16(s1i))), x1->m*y1->d);
-
 #if defined(__ARM_FEATURE_DOTPROD)
         // dot product into int32x4_t
         const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0ls), v0_0h, v1_0hs);
@@ -2637,24 +2672,26 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
 #endif
     }
 
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
 
+    float summs = 0;
+
     // Main loop
     for (int i = 0; i < nb; ++i) {
         const float * d0 = &x[i].d;
         const float * d1 = &y[i].d;
-        const float * m0 = &x[i].m;
+        //const float * m0 = &x[i].m;
+
+        summs += x[i].m * y[i].s;
 
         const __m256 d0v = _mm256_broadcast_ss( d0 );
         const __m256 d1v = _mm256_broadcast_ss( d1 );
-        const __m256 m0v = _mm256_broadcast_ss( m0 );
 
         // Compute combined scales
         const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
-        const __m256 d1m0 = _mm256_mul_ps( d1v, m0v );
 
         // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
         const __m256i bx = bytes_from_nibbles_32(x[i].qs);
@@ -2676,15 +2713,6 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
 
         // Accumulate d0*d1*x*y
         acc = _mm256_fmadd_ps( d0d1, xy, acc );
-
-        // Compute sum of y values
-        const __m256i y16_l = _mm256_cvtepi8_epi16( _mm256_castsi256_si128( by ) );
-        const __m256i y16_h = _mm256_cvtepi8_epi16( _mm256_extracti128_si256( by, 1 ) );
-        const __m256i ysumi = _mm256_madd_epi16( _mm256_add_epi16(y16_l, y16_h), ones );
-        const __m256 ysum = _mm256_cvtepi32_ps( ysumi );
-
-        // Accumulate d1*m0*y
-        acc = _mm256_fmadd_ps( d1m0, ysum, acc );
     }
 
     // Return horizontal sum of the acc vector
@@ -2693,7 +2721,7 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
     res = _mm_add_ps( res, _mm_movehl_ps( res, res ) );
     res = _mm_add_ss( res, _mm_movehdup_ps( res ) );
 
-    sumf = _mm_cvtss_f32( res );
+    sumf = _mm_cvtss_f32( res ) + summs;
 #else
     // scalar
     for (int i = 0; i < nb; i++) {

llama.cpp

@@ -1618,8 +1618,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         // quantize only 2D tensors
         quantize &= (tensor.ne.size() == 2);
 
-        // GG: uncomment this to keep the output layer in FP16
-        //if (tensor.name.rfind("output")) {
+        // uncomment this to keep the output layer in FP16
+        //if (tensor.name == "output.weight") {
         //    quantize = false;
         //}
 
@@ -2092,7 +2092,11 @@ void llama_set_kv_cache(
                          int   n_token_count) {
     // Make sure we have the same kv cache setup
     LLAMA_ASSERT(ctx->model.kv_self.buf.size == n_size);
+    void * k_data = ctx->model.kv_self.k->data; // remember data pointers
+    void * v_data = ctx->model.kv_self.v->data; // because their value is stored in buf and overwritten by memcpy
     memcpy(ctx->model.kv_self.buf.addr, kv_cache, n_size);
+    ctx->model.kv_self.k->data = k_data; // restore correct data pointers
+    ctx->model.kv_self.v->data = v_data;
     ctx->model.kv_self.n = n_token_count;
 }
 

pocs/vdot/CMakeLists.txt

@@ -2,3 +2,8 @@ set(TARGET vdot)
 add_executable(${TARGET} vdot.cpp)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET q8dot)
+add_executable(${TARGET} q8dot.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

pocs/vdot/q8dot.cpp

@@ -0,0 +1,172 @@
+#include <cstdio>
+#include <type_traits>
+#include <vector>
+#include <random>
+#include <chrono>
+#include <cstdlib>
+#include <cmath>
+#include <cassert>
+#include <cstring>
+#include <array>
+#include <type_traits>
+
+#include <ggml.h>
+
+constexpr int kVecSize = 1 << 16;
+
+// Copy-pasted from ggml.c
+#define QK4_0 32
+typedef struct {
+    float   d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    float   d;          // delta
+    float   m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+// Copy-pasted from ggml.c
+#define QK8_0 32
+typedef struct {
+    float   d;          // delta
+    float   s;          // d * sum(qs[i])
+    int8_t  qs[QK8_0];  // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == 2*sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+
+static_assert(QK4_1 == QK8_0, "QK4_1 and QK8_0 must be the same");
+static_assert(QK4_0 == QK8_0, "QK4_0 and QK8_0 must be the same");
+
+template <typename T>
+void fillQ4blocks(std::vector<T>& blocks, std::mt19937& rndm) {
+    for (auto& b : blocks) {
+        b.d = 1;
+        for (int i=0; i<QK4_1/2; ++i) {
+            uint8_t v1 = rndm() >> 28;
+            uint8_t v2 = rndm() >> 28;
+            b.qs[i] = v1 | (v2 << 4);
+        }
+    }
+}
+
+void fillQ80blocks(std::vector<block_q8_0>& blocks, std::mt19937& rndm) {
+    for (auto& b : blocks) {
+        b.d = 1;
+        int sum = 0;
+        for (int i=0; i<QK8_0; ++i) {
+            b.qs[i] = (rndm() >> 24) - 128;
+            sum += b.qs[i];
+        }
+        b.s = b.d * sum;
+    }
+}
+
+float simpleDot(const block_q4_0& x, const block_q8_0& y) {
+    int s1 = 0; //, s2 = 0;
+    for (int i=0; i<QK4_1/2; i+=2) {
+        int v1 = x.qs[i+0] & 0xf;
+        int v2 = x.qs[i+0] >> 4;
+        int v3 = x.qs[i+1] & 0xf;
+        int v4 = x.qs[i+1] >> 4;
+        int j = 2*i;
+        s1 += v1*y.qs[j] + v2*y.qs[j+1] + v3*y.qs[j+2] + v4*y.qs[j+3];
+        //s2 += y.qs[j] + y.qs[j+1] + y.qs[j+2] + y.qs[j+3];
+    }
+    return y.d * x.d * s1 - 8 * x.d * y.s;
+    //return y.d * x.d * (s1 - 8 * s2);
+}
+
+float simpleDot(const block_q4_1& x, const block_q8_0& y) {
+    int s1 = 0; //, s2 = 0;
+    for (int i=0; i<QK4_1/2; i+=2) {
+        int v1 = x.qs[i+0] & 0xf;
+        int v2 = x.qs[i+0] >> 4;
+        int v3 = x.qs[i+1] & 0xf;
+        int v4 = x.qs[i+1] >> 4;
+        int j = 2*i;
+        s1 += v1*y.qs[j] + v2*y.qs[j+1] + v3*y.qs[j+2] + v4*y.qs[j+3];
+        //s2 += y.qs[j] + y.qs[j+1] + y.qs[j+2] + y.qs[j+3];
+    }
+    return y.d * x.d * s1 + y.s * x.m;
+    //return y.d * (x.d * s1 + x.m * s2);
+}
+
+struct Stat {
+    double sum = 0, sumt = 0, sumt2 = 0, maxt = 0;
+    int nloop = 0;
+    void addResult(double s, double t) {
+        sum += s;
+        sumt += t; sumt2 += t*t; maxt = std::max(maxt, t);
+        ++nloop;
+    }
+    void reportResult(const char* title) const {
+        if (nloop < 1) {
+            printf("%s(%s): no result\n",__func__,title);
+            return;
+        }
+        printf("============ %s\n",title);
+        printf("<dot> = %g\n",sum/nloop);
+        auto t = sumt/nloop, dt = sumt2/nloop - t*t;
+        if (dt > 0) dt = sqrt(dt);
+        printf("<time> = %g +/- %g us. Max. time = %g us.\n",t,dt,maxt);
+    }
+};
+
+
+int main(int argc, char** argv) {
+
+    int nloop = argc > 1 ? atoi(argv[1]) : 10;
+    int type  = argc > 2 ? atoi(argv[2]) : 1;
+
+    std::mt19937 rndm(1234);
+
+    std::vector<block_q4_1> x41;
+    std::vector<block_q4_0> x40;
+    std::vector<block_q8_0> y(kVecSize);
+    if (type == 0) x40.resize(kVecSize);
+    else {
+        x41.resize(kVecSize);
+        for (auto& b : x41) b.m = 1;
+    }
+
+    auto ggml_type = type == 0 ? GGML_TYPE_Q4_0 : GGML_TYPE_Q4_1;
+
+    auto funcs = ggml_internal_get_quantize_fn(ggml_type);
+
+    Stat simple, ggml;
+
+    for (int iloop=0; iloop<nloop; ++iloop) {
+
+        if (type == 0) fillQ4blocks(x40, rndm);
+        else fillQ4blocks(x41, rndm);
+        fillQ80blocks(y, rndm);
+
+        auto t1 = std::chrono::high_resolution_clock::now();
+        double s = 0;
+        if (type == 0) for (int i=0; i<kVecSize; ++i) s += simpleDot(x40[i], y[i]);
+        else for (int i=0; i<kVecSize; ++i) s += simpleDot(x41[i], y[i]);
+        auto t2 = std::chrono::high_resolution_clock::now();
+        auto t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
+        if (iloop > 3) simple.addResult(s, t);
+
+        t1 = std::chrono::high_resolution_clock::now();
+        float fs;
+        if (type == 0) funcs.vec_dot_q(kVecSize * QK4_1, &fs, x40.data(), y.data());
+        else funcs.vec_dot_q(kVecSize * QK4_1, &fs, x41.data(), y.data());
+        t2 = std::chrono::high_resolution_clock::now();
+        t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
+        if (iloop > 3) ggml.addResult(fs, t);
+
+    }
+
+    // Report the time (and the average of the dot products so the compiler does not come up with the idea
+    // of optimizing away the function calls after figuring that the result is not used).
+    simple.reportResult("Simple");
+    ggml.reportResult("ggml");
+    return 0;
+}