Merge pull request #211 from jmorganca/update-llama-cpp

update llama.cpp

api/types.go

@@ -153,6 +153,7 @@ type Options struct {
 	NumCtx        int  `json:"num_ctx,omitempty"`
 	NumKeep       int  `json:"num_keep,omitempty"`
 	NumBatch      int  `json:"num_batch,omitempty"`
+	NumGQA        int  `json:"num_gqa,omitempty"`
 	NumGPU        int  `json:"num_gpu,omitempty"`
 	MainGPU       int  `json:"main_gpu,omitempty"`
 	LowVRAM       bool `json:"low_vram,omitempty"`
@@ -190,6 +191,7 @@ func DefaultOptions() Options {
 		NumCtx:   2048,
 		NumBatch: 1024,
 		NumGPU:   1,
+		NumGQA:   1,
 		LowVRAM:  false,
 		F16KV:    true,
 		UseMMap:  true,

llama/ggml-cuda.h

 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *

llama/ggml-metal.h

 //go:build darwin
 
 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *
@@ -89,6 +89,13 @@ void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor *
 // get data from the device into host memory
 void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
 
+// try to find operations that can be run concurrently in the graph
+// you should run it again if the topology of your graph changes
+void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+
+// if the graph has been optimized for concurrently dispatch
+bool ggml_metal_if_optimized(struct ggml_metal_context * ctx);
+
 // same as ggml_graph_compute but uses Metal
 // creates gf->n_threads command buffers in parallel
 void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);

llama/ggml-mpi.c

 //go:build mpi
 
 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *

llama/ggml-mpi.h

 //go:build mpi
 
 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *

llama/ggml-opencl.cpp

 //go:build opencl
 
 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *

llama/ggml-opencl.h

 //go:build opencl
 
 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *

llama/ggml.h

 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *
@@ -225,6 +225,7 @@
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_SRC           6
 #define GGML_MAX_NAME          48
+#define GGML_MAX_OP_PARAMS     32
 #define GGML_DEFAULT_N_THREADS 4
 
 
@@ -233,6 +234,7 @@
 
 #define GGML_UNUSED(x) (void)(x)
 
+#define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
 #define GGML_ASSERT(x) \
     do { \
@@ -355,16 +357,6 @@ extern "C" {
         GGML_OP_ARGMAX,
         GGML_OP_REPEAT,
         GGML_OP_REPEAT_BACK,
-        GGML_OP_ABS,
-        GGML_OP_SGN,
-        GGML_OP_NEG,
-        GGML_OP_STEP,
-        GGML_OP_TANH,
-        GGML_OP_ELU,
-        GGML_OP_RELU,
-        GGML_OP_GELU,
-        GGML_OP_GELU_QUICK,
-        GGML_OP_SILU,
         GGML_OP_SILU_BACK,
         GGML_OP_NORM, // normalize
         GGML_OP_RMS_NORM,
@@ -403,6 +395,8 @@ extern "C" {
         GGML_OP_WIN_PART,
         GGML_OP_WIN_UNPART,
 
+        GGML_OP_UNARY,
+
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
@@ -416,6 +410,24 @@ extern "C" {
         GGML_OP_COUNT,
     };
 
+    enum ggml_unary_op {
+        GGML_UNARY_OP_ABS,
+        GGML_UNARY_OP_SGN,
+        GGML_UNARY_OP_NEG,
+        GGML_UNARY_OP_STEP,
+        GGML_UNARY_OP_TANH,
+        GGML_UNARY_OP_ELU,
+        GGML_UNARY_OP_RELU,
+        GGML_UNARY_OP_GELU,
+        GGML_UNARY_OP_GELU_QUICK,
+        GGML_UNARY_OP_SILU,
+    };
+
+    enum ggml_object_type {
+        GGML_OBJECT_TENSOR,
+        GGML_OBJECT_GRAPH,
+        GGML_OBJECT_WORK_BUFFER
+    };
 
     // ggml object
     struct ggml_object {
@@ -424,7 +436,9 @@ extern "C" {
 
         struct ggml_object * next;
 
-        char padding[8];
+        enum ggml_object_type type;
+
+        char padding[4];
     };
 
     static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
@@ -444,6 +458,9 @@ extern "C" {
         // compute data
         enum ggml_op op;
 
+        // op params - allocated as int32_t for alignment
+        int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+
         bool is_param;
 
         struct ggml_tensor * grad;
@@ -460,7 +477,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[8];
+        char padding[4];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -481,6 +498,11 @@ extern "C" {
         void * abort_callback_data;
     };
 
+    // next prime after GGML_MAX_NODES
+    // #define GGML_GRAPH_HASHTABLE_SIZE 4099
+    // next prime after GGML_MAX_NODES * 2 (nodes + leafs)
+    #define GGML_GRAPH_HASHTABLE_SIZE 8273
+
     // computation graph
     struct ggml_cgraph {
         int n_nodes;
@@ -490,12 +512,16 @@ extern "C" {
         struct ggml_tensor * grads[GGML_MAX_NODES];
         struct ggml_tensor * leafs[GGML_MAX_NODES];
 
+        void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE];
+
         // performance
         int     perf_runs;
         int64_t perf_cycles;
         int64_t perf_time_us;
     };
 
+    static const size_t GGML_GRAPH_SIZE = sizeof(struct ggml_cgraph);
+
     // scratch buffer
     struct ggml_scratch {
         size_t offs;
@@ -557,6 +583,7 @@ extern "C" {
 
     GGML_API const char * ggml_type_name(enum ggml_type type);
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
+    GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
@@ -580,6 +607,7 @@ extern "C" {
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 
     GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
+    GGML_API bool    ggml_get_no_alloc(struct ggml_context * ctx);
     GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
 
     GGML_API void *  ggml_get_mem_buffer     (const struct ggml_context * ctx);
@@ -639,9 +667,11 @@ extern "C" {
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
-    GGML_API const char *         ggml_get_name(const struct ggml_tensor * tensor);
-    GGML_API struct ggml_tensor * ggml_set_name(struct ggml_tensor * tensor, const char * name);
-    GGML_API struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char * fmt, ...);
+    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
+
+    GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
+    GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);
 
     //
     // operations on tensors with backpropagation
@@ -651,6 +681,11 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_dup_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_add(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -875,14 +910,17 @@ extern "C" {
 
     GGML_API struct ggml_tensor * ggml_rms_norm(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_rms_norm_inplace(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     // a - x
     // b - dy
+    // TODO: update with configurable eps
     GGML_API struct ggml_tensor * ggml_rms_norm_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -974,11 +1012,22 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // a -> b, in-place, return view(b)
+    GGML_API struct ggml_tensor * ggml_cpy_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // make contiguous
     GGML_API struct ggml_tensor * ggml_cont(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // make contiguous, in-place
+    GGML_API struct ggml_tensor * ggml_cont_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // return view(a), b specifies the new shape
     // TODO: when we start computing gradient, make a copy instead of view
     GGML_API struct ggml_tensor * ggml_reshape(
@@ -1154,9 +1203,9 @@ 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
@@ -1165,7 +1214,8 @@ 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)
@@ -1289,6 +1339,16 @@ extern "C" {
     typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
     typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
 
+    GGML_API struct ggml_tensor * ggml_unary(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             enum ggml_unary_op op);
+
+    GGML_API struct ggml_tensor * ggml_unary_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        enum ggml_unary_op op);
+
     GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
@@ -1368,11 +1428,17 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * tensor);
 
+
     GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
 
     GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
     GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);
 
+    // graph allocation in a context
+    GGML_API struct ggml_cgraph * ggml_new_graph        (struct ggml_context * ctx);
+    GGML_API struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API size_t ggml_graph_overhead(void);
+
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
     GGML_API struct ggml_cplan ggml_graph_plan   (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);

llama/k_quants.c

 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *
@@ -1692,6 +1692,62 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) + summs;
 
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint32_t ud, um;
+    const uint8_t * restrict db = (const uint8_t *)&ud;
+    const uint8_t * restrict mb = (const uint8_t *)&um;
+
+    float summs = 0;
+
+    // TODO: optimize this
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
+        ud = (sc[0] >> 0) & 0x0f0f0f0f;
+        um = (sc[0] >> 4) & 0x0f0f0f0f;
+
+        int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3];
+        summs += dmin * smin;
+
+        const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
+        const __m128i q2_0 = _mm_and_si128(q2bits, m3);
+        const __m128i q2_1 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+        const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+        const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p0 = _mm_maddubs_epi16(q2_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i p1 = _mm_maddubs_epi16(q2_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i p2 = _mm_maddubs_epi16(q2_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i p3 = _mm_maddubs_epi16(q2_3, _mm256_extractf128_si256(q8_1, 1));
+
+        const __m256i p_0 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
+        const __m256i p_1 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
+        const __m256i p_2 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
+        const __m256i p_3 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3)), acc);
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
 #else
 
     float sumf = 0;
@@ -2321,6 +2377,93 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint64_t aux64;
+
+    uint16_t aux16[2];
+    const int8_t * aux8 = (const int8_t *)aux16;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint16_t a = *(const uint16_t *)x[i].scales;
+        aux16[0] = a & 0x0f0f;
+        aux16[1] = (a >> 4) & 0x0f0f;
+
+        const __m128i scale_0 = _mm_set1_epi16(aux8[0] - 8);
+        const __m128i scale_1 = _mm_set1_epi16(aux8[2] - 8);
+        const __m128i scale_2 = _mm_set1_epi16(aux8[1] - 8);
+        const __m128i scale_3 = _mm_set1_epi16(aux8[3] - 8);
+
+        memcpy(&aux64, x[i].hmask, 8);
+
+        __m128i q3h_0 = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
+        __m128i q3h_1 = _mm_srli_epi16(q3h_0, 2);
+        __m128i q3h_2 = _mm_srli_epi16(q3h_0, 4);
+        __m128i q3h_3 = _mm_srli_epi16(q3h_0, 6);
+        q3h_0 = _mm_slli_epi16(_mm_andnot_si128(q3h_0, m1), 2);
+        q3h_1 = _mm_slli_epi16(_mm_andnot_si128(q3h_1, m1), 2);
+        q3h_2 = _mm_slli_epi16(_mm_andnot_si128(q3h_2, m1), 2);
+        q3h_3 = _mm_slli_epi16(_mm_andnot_si128(q3h_3, m1), 2);
+
+        // load low 2 bits
+        const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
+
+        // prepare low and high bits
+        const __m128i q3l_0 = _mm_and_si128(q3bits, m3);
+        const __m128i q3l_1 = _mm_and_si128(_mm_srli_epi16(q3bits, 2), m3);
+        const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits, 4), m3);
+        const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits, 6), m3);
+
+        // load Q8 quants
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm_maddubs_epi16,
+        // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+        // and 2 if the high bit was set)
+        const __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, _mm256_extractf128_si256(q8_1, 1));
+
+        __m128i p16_0 = _mm_maddubs_epi16(q3l_0, _mm256_extractf128_si256(q8_0, 0));
+        __m128i p16_1 = _mm_maddubs_epi16(q3l_1, _mm256_extractf128_si256(q8_0, 1));
+        __m128i p16_2 = _mm_maddubs_epi16(q3l_2, _mm256_extractf128_si256(q8_1, 0));
+        __m128i p16_3 = _mm_maddubs_epi16(q3l_3, _mm256_extractf128_si256(q8_1, 1));
+
+        p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+        p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+        p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+        p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+
+        // multiply with scales
+        p16_0 = _mm_madd_epi16(scale_0, p16_0);
+        p16_1 = _mm_madd_epi16(scale_1, p16_1);
+        p16_2 = _mm_madd_epi16(scale_2, p16_2);
+        p16_3 = _mm_madd_epi16(scale_3, p16_3);
+
+        p16_0 = _mm_add_epi32(p16_0, p16_2);
+        p16_1 = _mm_add_epi32(p16_1, p16_3);
+        __m256i p16 = _mm256_set_m128i(p16_1, p16_0);
+
+        // multiply with block scale and accumulate
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16)), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     int8_t  aux8[QK_K];
@@ -2807,6 +2950,60 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) - summs;
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0;
+
+    uint16_t aux16[2];
+    const uint8_t * scales = (const uint8_t *)aux16;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = ggml_fp16_to_fp32(x[i].d[0]) * y[i].d;
+        const float m = ggml_fp16_to_fp32(x[i].d[1]) * y[i].d;
+        const __m256 vd = _mm256_set1_ps(d);
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux16[0] = a[0] & 0x0f0f;
+        aux16[1] = (a[0] >> 4) & 0x0f0f;
+
+        summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4);
+        const __m128i q4bits_0 = _mm256_extractf128_si256(q4bits, 0);
+        const __m128i q4bits_1 = _mm256_extractf128_si256(q4bits, 1);
+        const __m128i q4_0 = _mm_and_si128(q4bits_0, m4);
+        const __m128i q4_1 = _mm_and_si128(q4bits_1, m4);
+        const __m128i q4_2 = _mm_and_si128(_mm_srli_epi16(q4bits_0, 4), m4);
+        const __m128i q4_3 = _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
+
+        const __m128i p32_0 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_0);
+        const __m128i p32_1 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_1);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_1, p32_0))), acc);
+
+        const __m128i p32_2 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_2);
+        const __m128i p32_3 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_3);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_3, p32_2))), acc);
+
+    }
+
+    *s = hsum_float_8(acc) - summs;
+
 #else
 
     uint8_t aux8[QK_K];
@@ -3321,10 +3518,66 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
-#else
+#elif defined __AVX__
 
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i mone  = _mm_set1_epi8(1);
 
-    uint8_t aux8[QK_K];
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
+
+        const __m128i scale_0 = _mm_set1_epi16(x[i].scales[0]);
+        const __m128i scale_1 = _mm_set1_epi16(x[i].scales[1]);
+        const __m128i scale_2 = _mm_set1_epi16(x[i].scales[2]);
+        const __m128i scale_3 = _mm_set1_epi16(x[i].scales[3]);
+
+        int64_t aux64;
+        memcpy(&aux64, x[i].qh, 8);
+        const __m128i haux128_0 = _mm_set_epi64x(aux64 >> 1, aux64);
+        const __m128i haux128_1 = _mm_srli_epi16(haux128_0, 2);
+
+        const __m128i q5h_0 = _mm_slli_epi16(_mm_andnot_si128(haux128_0, mone), 4);
+        const __m128i q5h_1 = _mm_slli_epi16(_mm_andnot_si128(haux128_1, mone), 4);
+        const __m128i q5h_2 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_0, 4), mone), 4);
+        const __m128i q5h_3 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_1, 4), mone), 4);
+
+        const __m128i q5l_0 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 0), m4);
+        const __m128i q5l_1 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 1), m4);
+        const __m128i q5l_2 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 0), 4), m4);
+        const __m128i q5l_3 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 1), 4), m4);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5l_0, _mm256_extractf128_si256(q8_0, 0)));
+        const __m128i p16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5l_1, _mm256_extractf128_si256(q8_0, 1)));
+        const __m128i p16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5l_2, _mm256_extractf128_si256(q8_1, 0)));
+        const __m128i p16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5l_3, _mm256_extractf128_si256(q8_1, 1)));
+        const __m128i s16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5h_0, _mm256_extractf128_si256(q8_0, 0)));
+        const __m128i s16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5h_1, _mm256_extractf128_si256(q8_0, 1)));
+        const __m128i s16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5h_2, _mm256_extractf128_si256(q8_1, 0)));
+        const __m128i s16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5h_3, _mm256_extractf128_si256(q8_1, 1)));
+
+        const __m128i dot_0 = _mm_sub_epi32(_mm_add_epi32(p16_0, p16_2), _mm_add_epi32(s16_0, s16_2));
+        const __m128i dot_1 = _mm_sub_epi32(_mm_add_epi32(p16_1, p16_3), _mm_add_epi32(s16_1, s16_3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_set_m128i(dot_1, dot_0))), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    int8_t aux8[QK_K];
     int16_t aux16[16];
     float   sums [8];
     memset(sums, 0, 8*sizeof(float));
@@ -3334,7 +3587,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         const uint8_t * restrict q4 = x[i].qs;
         const uint8_t * restrict hm = x[i].qh;
         const  int8_t * restrict q8 = y[i].qs;
-        uint8_t * restrict a = aux8;
+        int8_t * restrict a = aux8;
         for (int l = 0; l < 32; ++l) {
             a[l+ 0] = q4[l] & 0xF;
             a[l+32] = q4[l]  >> 4;
@@ -3884,6 +4137,77 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(3);
+    const __m128i m32s = _mm_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]);
+        const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]);
+        const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]);
+        const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1);
+        const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3);
+
+        const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
+        const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
+
+        const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH, m2), 4);
+        const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 2), m2), 4);
+        const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 4), m2), 4);
+        const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 6), m2), 4);
+
+        const __m128i q4_0 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 0), m4), q4h_0);
+        const __m128i q4_1 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 1), m4), q4h_1);
+        const __m128i q4_2 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 0), 4), m4), q4h_2);
+        const __m128i q4_3 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 1), 4), m4), q4h_3);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        __m128i q8s_0 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 0));
+        __m128i q8s_1 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 1));
+        __m128i q8s_2 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 0));
+        __m128i q8s_3 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 1));
+
+        __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
+        __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
+        __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
+        __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
+
+        p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+        p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+        p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+        p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+
+        p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
+        p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1);
+        p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
+        p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3);
+
+        sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+        sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(_mm256_set_m128i(sumi_1, sumi_0))), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     int8_t  aux8[QK_K];

llama/k_quants.h

 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *

llama/llama-util.h

 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *

llama/llama.go

@@ -127,6 +127,7 @@ func New(model string, opts api.Options) (*LLM, error) {
 	params.seed = C.uint(llm.Seed)
 	params.n_ctx = C.int(llm.NumCtx)
 	params.n_batch = C.int(llm.NumBatch)
+	params.n_gqa = C.int(llm.NumGQA)
 	params.n_gpu_layers = C.int(llm.NumGPU)
 	params.main_gpu = C.int(llm.MainGPU)
 	params.low_vram = C.bool(llm.LowVRAM)

llama/llama.h

 /**
- * llama.cpp - git e782c9e735f93ab4767ffc37462c523b73a17ddc
+ * llama.cpp - git 7c529cede6e84054e77a3eceab31c53de7b2f55b
  *
  * MIT License
  *
@@ -79,6 +79,10 @@
 #define LLAMA_SUPPORTS_GPU_OFFLOAD
 #endif
 
+#ifndef LLAMA_DEFAULT_RMS_EPS
+#define LLAMA_DEFAULT_RMS_EPS 5e-6f
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -109,12 +113,15 @@ extern "C" {
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
    struct llama_context_params {
-        uint32_t seed;                         // RNG seed, -1 for random
-        int32_t  n_ctx;                        // text context
-        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
+        uint32_t seed;         // RNG seed, -1 for random
+        int32_t  n_ctx;        // text context
+        int32_t  n_batch;      // prompt processing batch size
+        int32_t  n_gqa;        // grouped-query attention (TEMP - will be moved to model hparams)
+        float    rms_norm_eps; // rms norm epsilon (TEMP - will be moved to model hparams)
+        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
+
+        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
@@ -165,6 +172,40 @@ extern "C" {
         bool quantize_output_tensor; // quantize output.weight
     } llama_model_quantize_params;
 
+    // grammar types
+    struct llama_grammar;
+
+    // grammar element type
+    enum llama_gretype {
+        // end of rule definition
+        LLAMA_GRETYPE_END            = 0,
+
+        // start of alternate definition for rule
+        LLAMA_GRETYPE_ALT            = 1,
+
+        // non-terminal element: reference to rule
+        LLAMA_GRETYPE_RULE_REF       = 2,
+
+        // terminal element: character (code point)
+        LLAMA_GRETYPE_CHAR           = 3,
+
+        // inverse char(s) ([^a], [^a-b] [^abc])
+        LLAMA_GRETYPE_CHAR_NOT       = 4,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
+        // be an inclusive range ([a-z])
+        LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or
+        // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
+        LLAMA_GRETYPE_CHAR_ALT       = 6,
+    };
+
+    typedef struct llama_grammar_element {
+        enum llama_gretype type;
+        uint32_t           value; // Unicode code point or rule ID
+    } llama_grammar_element;
+
     // performance timing information
     struct llama_timings {
         double t_start_ms;
@@ -357,6 +398,15 @@ extern "C" {
     LLAMA_API llama_token llama_token_eos();  // end-of-sentence
     LLAMA_API llama_token llama_token_nl();   // next-line
 
+    // Grammar
+    //
+    LLAMA_API struct llama_grammar * llama_grammar_init(
+            const llama_grammar_element ** rules,
+                                 size_t    n_rules,
+                                 size_t    start_rule_index);
+
+    LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
+
     // Sampling functions
 
     /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
@@ -369,13 +419,11 @@ 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);
@@ -393,6 +441,9 @@ extern "C" {
     LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
     LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
 
+    /// @details Apply constraints from grammar
+    LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
+
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@@ -414,6 +465,9 @@ extern "C" {
     /// @details Randomly selects a token from the candidates based on their probabilities.
     LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
 
+    /// @details Accepts the sampled token into the grammar
+    LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
+
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
     LLAMA_API void llama_print_timings(struct llama_context * ctx);