llama_cpp 0.3.1 → 0.3.2

data/ext/llama_cpp/src/ggml.h

@@ -201,6 +201,8 @@
 #define GGML_MAX_NAME          48
 #define GGML_DEFAULT_N_THREADS 4
 
+#define GGML_UNUSED(x) (void)(x)
+
 #define GGML_ASSERT(x) \
     do { \
         if (!(x)) { \
@@ -209,6 +211,30 @@
         } \
     } while (0)
 
+// used to copy the number of elements and stride in bytes of tensors into local variables.
+// main purpose is to reduce code duplication and improve readability.
+//
+// example:
+//
+//    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
+//    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb);
+//
+#define GGML_TENSOR_LOCALS_1(type, prefix, pointer, array) \
+    const type prefix##0 = (pointer)->array[0]; \
+    GGML_UNUSED(prefix##0);
+#define GGML_TENSOR_LOCALS_2(type, prefix, pointer, array) \
+    GGML_TENSOR_LOCALS_1    (type, prefix, pointer, array) \
+    const type prefix##1 = (pointer)->array[1]; \
+    GGML_UNUSED(prefix##1);
+#define GGML_TENSOR_LOCALS_3(type, prefix, pointer, array) \
+    GGML_TENSOR_LOCALS_2    (type, prefix, pointer, array) \
+    const type prefix##2 = (pointer)->array[2]; \
+    GGML_UNUSED(prefix##2);
+#define GGML_TENSOR_LOCALS(type, prefix, pointer, array) \
+    GGML_TENSOR_LOCALS_3  (type, prefix, pointer, array) \
+    const type prefix##3 = (pointer)->array[3]; \
+    GGML_UNUSED(prefix##3);
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -224,8 +250,8 @@ extern "C" {
     GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t x);
     GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x);
 
-    GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, size_t n);
-    GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, size_t n);
+    GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n);
+    GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n);
 
     struct ggml_object;
     struct ggml_context;
@@ -295,12 +321,15 @@ extern "C" {
         GGML_OP_SUM,
         GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
+        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,
@@ -332,9 +361,8 @@ extern "C" {
         GGML_OP_ROPE_BACK,
         GGML_OP_ALIBI,
         GGML_OP_CLAMP,
-        GGML_OP_CONV_1D_S1_PH,
-        GGML_OP_CONV_1D_S2_PH,
-        GGML_OP_CONV_2D_SK_P0,
+        GGML_OP_CONV_1D,
+        GGML_OP_CONV_2D,
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
@@ -444,6 +472,9 @@ extern "C" {
 
 
     // compute types
+
+    // NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled.
+    // This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995.
     enum ggml_task_type {
         GGML_TASK_INIT = 0,
         GGML_TASK_COMPUTE,
@@ -687,6 +718,11 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // argmax along rows
+    GGML_API struct ggml_tensor * ggml_argmax(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // if a is the same shape as b, and a is not parameter, return a
     // otherwise, return a new tensor: repeat(a) to fit in b
     GGML_API struct ggml_tensor * ggml_repeat(
@@ -731,6 +767,22 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_tanh(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_tanh_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_elu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_elu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_relu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1081,58 +1133,33 @@ extern "C" {
             float                 min,
             float                 max);
 
-    // TODO: implement general-purpose convolutions
-    // GGML_API struct ggml_tensor * ggml_conv_1d(
-    //        struct ggml_context * ctx,
-    //        struct ggml_tensor  * a,
-    //        struct ggml_tensor  * b,
-    //        int                   s0
-    //        int                   p0,
-    //        int                   d0);
-    //
-    // GGML_API struct ggml_tensor * ggml_conv_2d(
-    //        struct ggml_context * ctx,
-    //        struct ggml_tensor  * a,
-    //        struct ggml_tensor  * b,
-    //        int                   s0,
-    //        int                   s1,
-    //        int                   p0,
-    //        int                   p1,
-    //        int                   d0,
-    //        int                   d1);
-
-    // padding = half
-    // TODO: we don't support extra parameters for now
-    //       that's why we are hard-coding the stride, padding, and dilation
-    //       not great ..
-    // example:
-    // a:      3   80  768    1
-    // b:   3000   80    1    1
-    // res: 3000  768    1    1
-    // used in whisper
-    GGML_API struct ggml_tensor * ggml_conv_1d_s1_ph(
+    GGML_API struct ggml_tensor * ggml_conv_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
 
-    // used in whisper
-    GGML_API struct ggml_tensor * ggml_conv_1d_s2_ph(
+    GGML_API struct ggml_tensor * ggml_conv_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            int                   s0,
+            int                   s1,
+            int                   p0,
+            int                   p1,
+            int                   d0,
+            int                   d1);
 
-    // kernel size is a->ne[0] x a->ne[1]
-    // stride is equal to kernel size
-    // padding is zero
-    // example:
-    // a:     16   16    3  768
-    // b:   1024 1024    3    1
-    // res:   64   64  768    1
-    // used in sam
-    GGML_API struct ggml_tensor * ggml_conv_2d_sk_p0(
+    // conv_1d with padding = half
+    // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
+    GGML_API struct ggml_tensor* ggml_conv_1d_ph(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            int                   s,
+            int                   d);
 
     GGML_API struct ggml_tensor * ggml_flash_attn(
             struct ggml_context * ctx,
@@ -1488,25 +1515,24 @@ extern "C" {
     //
 
 #ifdef  __cplusplus
-    // restrict not standard in C++
+// restrict not standard in C++
 #define GGML_RESTRICT
 #else
 #define GGML_RESTRICT restrict
 #endif
-    typedef void (*dequantize_row_q_t)(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-    typedef void (*quantize_row_q_t)  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-    typedef void (*vec_dot_q_t)       (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
+    typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+    typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int k);
+    typedef void (*ggml_vec_dot_t)   (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
 
     typedef struct {
-        dequantize_row_q_t dequantize_row_q;
-        quantize_row_q_t   quantize_row_q;
-        quantize_row_q_t   quantize_row_q_reference;
-        quantize_row_q_t   quantize_row_q_dot;
-        vec_dot_q_t        vec_dot_q;
-        enum ggml_type     vec_dot_type;
-    } quantize_fns_t;
-
-    quantize_fns_t ggml_internal_get_quantize_fn(size_t i);
+        ggml_to_float_t   to_float;
+        ggml_from_float_t from_float;
+        ggml_from_float_t from_float_reference;
+        ggml_vec_dot_t    vec_dot;
+        enum ggml_type    vec_dot_type;
+    } ggml_type_traits_t;
+
+    ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type i);
 
 #ifdef  __cplusplus
 }

data/ext/llama_cpp/src/llama.cpp

@@ -66,6 +66,7 @@ enum e_model {
     MODEL_65B,
 };
 
+static const size_t kB = 1024;
 static const size_t MB = 1024*1024;
 
 // computed for n_ctx == 2048
@@ -129,6 +130,34 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
     return k_sizes;
 }
 
+// amount of VRAM needed per batch size to hold temporary results
+// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
+{
+    static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,   512ull * kB },
+        { MODEL_7B,   512ull * kB },
+        { MODEL_13B,  640ull * kB },
+        { MODEL_30B,  768ull * kB },
+        { MODEL_65B, 1536ull * kB },
+    };
+    return k_sizes;
+}
+
+// amount of VRAM needed per batch size and context to hold temporary results
+// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
+{
+    static std::map<e_model, size_t> k_sizes = {
+        { MODEL_3B,  128ull },
+        { MODEL_7B,  128ull },
+        { MODEL_13B, 160ull },
+        { MODEL_30B, 208ull },
+        { MODEL_65B, 416ull },
+    };
+    return k_sizes;
+}
+
 // default hparams (LLaMA 7B)
 struct llama_hparams {
     uint32_t n_vocab = 32000;
@@ -165,8 +194,8 @@ struct llama_layer {
 };
 
 struct llama_kv_cache {
-    struct ggml_tensor * k;
-    struct ggml_tensor * v;
+    struct ggml_tensor * k = NULL;
+    struct ggml_tensor * v = NULL;
 
     struct ggml_context * ctx = NULL;
 
@@ -253,7 +282,13 @@ struct llama_model {
 
 struct llama_context {
     llama_context(const llama_model & model, const llama_vocab & vocab) : model(model), vocab(vocab), t_load_us(model.t_load_us), t_start_us(model.t_start_us) {}
-
+#ifdef GGML_USE_METAL
+    ~llama_context() {
+        if (ctx_metal) {
+            ggml_metal_free(ctx_metal);
+        }
+    }
+#endif
     std::mt19937 rng;
 
     bool has_evaluated_once = false;
@@ -446,9 +481,7 @@ struct llama_file_loader {
             std::string word = file.read_string(len);
 
             float score = 0.0f;
-            if (file_version >= LLAMA_FILE_VERSION_GGMF_V1) {
-                file.read_raw(&score, sizeof(score));
-            }
+            file.read_raw(&score, sizeof(score));
 
             vocab.token_to_id[word] = i;
 
@@ -1112,14 +1145,18 @@ static void llama_model_load_internal(
             fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
             ggml_cuda_set_scratch_size(0); // disable scratch
         } else {
-            vram_scratch = n_batch * MB;
+            const size_t vram_scratch_base = VRAM_REQ_SCRATCH_BASE().at(model.type);
+            const size_t vram_scratch_per_context = VRAM_REQ_SCRATCH_PER_CONTEXT().at(model.type);
+            vram_scratch = n_batch * (vram_scratch_base + n_ctx * vram_scratch_per_context);
             ggml_cuda_set_scratch_size(vram_scratch);
             if (n_gpu_layers > 0) {
-                fprintf(stderr, "%s: allocating batch_size x 1 MB = %zd MB VRAM for the scratch buffer\n",
-                        __func__, vram_scratch / MB);
+                fprintf(stderr, "%s: allocating batch_size x (%zd kB + n_ctx x %zd B) = %zd MB VRAM for the scratch buffer\n",
+                        __func__, vram_scratch_base / kB, vram_scratch_per_context,
+                        (vram_scratch + MB - 1) / MB); // round up
             }
         }
 #endif // GGML_USE_CUBLAS
+
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
 
@@ -1128,6 +1165,10 @@ static void llama_model_load_internal(
             fprintf(stderr, "%s: offloading non-repeating layers to GPU\n", __func__);
         }
         size_t vram_kv_cache = 0;
+
+#ifdef GGML_USE_CUBLAS
+        const int max_backend_supported_layers = hparams.n_layer + 3;
+        const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
         if (n_gpu_layers > (int) hparams.n_layer + 1) {
             if (low_vram) {
                 fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
@@ -1144,14 +1185,18 @@ static void llama_model_load_internal(
                 vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
             }
         }
-        const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
+#elif defined(GGML_USE_CLBLAST)
+        const int max_backend_supported_layers = hparams.n_layer + 1;
+        const int max_offloadable_layers = hparams.n_layer + 1;
+#endif // GGML_USE_CUBLAS
+
         fprintf(stderr, "%s: offloaded %d/%d layers to GPU\n",
-                __func__, std::min(n_gpu_layers, max_offloadable_layers), hparams.n_layer + 3);
+                __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers);
         fprintf(stderr, "%s: total VRAM used: %zu MB\n",
                 __func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up
 #else
         (void) n_gpu_layers;
-#endif
+#endif // defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
     }
 
     // populate `tensors_by_name`
@@ -1860,10 +1905,10 @@ void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * can
         return;
     }
 
-    const int64_t t_start_sample_us = ggml_time_us();
-
     llama_sample_softmax(ctx, candidates);
 
+    const int64_t t_start_sample_us = ggml_time_us();
+
     // Compute the cumulative probabilities
     float cum_sum = 0.0f;
     size_t last_idx = candidates->size;
@@ -1892,9 +1937,8 @@ void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array *
         return;
     }
 
-    const int64_t t_start_sample_us = ggml_time_us();
-
     llama_sample_softmax(nullptr, candidates);
+    const int64_t t_start_sample_us = ggml_time_us();
 
     // Compute the first and second derivatives
     std::vector<float> first_derivatives(candidates->size - 1);
@@ -1946,11 +1990,11 @@ void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * c
         return;
     }
 
-    const int64_t t_start_sample_us = ggml_time_us();
-
     // Compute the softmax of logits and calculate entropy
     llama_sample_softmax(nullptr, candidates);
 
+    const int64_t t_start_sample_us = ggml_time_us();
+
     float entropy = 0.0f;
     for (size_t i = 0; i < candidates->size; ++i) {
         entropy += -candidates->data[i].p * logf(candidates->data[i].p);
@@ -2119,13 +2163,11 @@ llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_
 
     if (ctx) {
         ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
-        ctx->n_sample++;
     }
     return X;
 }
 
 llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu) {
-    assert(ctx);
     int64_t t_start_sample_us;
     t_start_sample_us = ggml_time_us();
 
@@ -2140,13 +2182,14 @@ llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_tok
         candidates->size = 1;
     }
 
+    if (ctx) {
+        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
+    }
+
     // Normalize the probabilities of the remaining words
     llama_sample_softmax(ctx, candidates);
 
     // Sample the next word X from the remaining words
-    if (ctx) {
-        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
-    }
     llama_token X = llama_sample_token(ctx, candidates);
     t_start_sample_us = ggml_time_us();
 
@@ -2214,10 +2257,10 @@ static void llama_convert_tensor_internal(const llama_load_tensor & tensor, llam
     }
     float * f32_output = (float *) output.addr;
 
-    quantize_fns_t qtype;
+    ggml_type_traits_t qtype;
     if (ggml_is_quantized(tensor.type)) {
-        qtype = ggml_internal_get_quantize_fn(tensor.type);
-        if (qtype.dequantize_row_q == NULL) {
+        qtype = ggml_internal_get_type_traits(tensor.type);
+        if (qtype.to_float == NULL) {
             throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor.type)));
         }
     } else if (tensor.type != GGML_TYPE_F16) {
@@ -2228,7 +2271,7 @@ static void llama_convert_tensor_internal(const llama_load_tensor & tensor, llam
         if (tensor.type == GGML_TYPE_F16) {
             ggml_fp16_to_fp32_row((ggml_fp16_t *)tensor.data, f32_output, nelements);
         } else if (ggml_is_quantized(tensor.type)) {
-            qtype.dequantize_row_q(tensor.data, f32_output, nelements);
+            qtype.to_float(tensor.data, f32_output, nelements);
         } else {
             LLAMA_ASSERT(false); // unreachable
         }
@@ -2253,7 +2296,7 @@ static void llama_convert_tensor_internal(const llama_load_tensor & tensor, llam
             if (typ == GGML_TYPE_F16) {
                 ggml_fp16_to_fp32_row((ggml_fp16_t *)inbuf, outbuf, nels);
             } else {
-                qtype.dequantize_row_q(inbuf, outbuf, nels);
+                qtype.to_float(inbuf, outbuf, nels);
             }
         };
         workers.push_back(std::thread(compute, tensor.type, tensor.data + in_buff_offs, f32_output + out_buff_offs, thr_elems));
@@ -3219,7 +3262,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
     return nread;
 }
 
-bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+static bool llama_load_session_file_internal(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
     llama_file file(path_session, "rb");
 
     // sanity checks
@@ -3273,6 +3316,15 @@ bool llama_load_session_file(struct llama_context * ctx, const char * path_sessi
     return true;
 }
 
+bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+    try {
+        return llama_load_session_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out);
+    } catch (const std::exception & err) {
+        fprintf(stderr, "error loading session file: %s\n", err.what());
+        return false;
+    }
+}
+
 bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
     llama_file file(path_session, "wb");
 
@@ -3428,23 +3480,35 @@ llama_token llama_token_nl() {
     return 13;
 }
 
+struct llama_timings llama_get_timings(struct llama_context * ctx) {
+    struct llama_timings result = {
+        /*.t_start_ms  =*/ 1e-3 * ctx->t_start_us,
+        /*.t_end_ms    =*/ 1.00 * ggml_time_ms(),
+        /*.t_load_ms   =*/ 1e-3 * ctx->t_load_us,
+        /*.t_sample_ms =*/ 1e-3 * ctx->t_sample_us,
+        /*.t_p_eval_ms =*/ 1e-3 * ctx->t_p_eval_us,
+        /*.t_eval_ms   =*/ 1e-3 * ctx->t_eval_us,
 
-void llama_print_timings(struct llama_context * ctx) {
-    const int64_t t_end_us = ggml_time_us();
+        /*.n_sample =*/ std::max(1, ctx->n_sample),
+        /*.n_p_eval =*/ std::max(1, ctx->n_p_eval),
+        /*.n_eval   =*/ std::max(1, ctx->n_eval),
+    };
 
-    const int32_t n_sample = std::max(1, ctx->n_sample);
-    const int32_t n_eval   = std::max(1, ctx->n_eval);
-    const int32_t n_p_eval = std::max(1, ctx->n_p_eval);
+    return result;
+}
+
+void llama_print_timings(struct llama_context * ctx) {
+    const llama_timings timings = llama_get_timings(ctx);
 
     fprintf(stderr, "\n");
-    fprintf(stderr, "%s:        load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0);
+    fprintf(stderr, "%s:        load time = %8.2f ms\n", __func__, timings.t_load_ms);
     fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, 1e-3 * ctx->t_sample_us, n_sample, 1e-3 * ctx->t_sample_us / n_sample, 1e6 / ctx->t_sample_us * n_sample);
+            __func__, timings.t_sample_ms, timings.n_sample, timings.t_sample_ms / timings.n_sample, 1e3 / timings.t_sample_ms * timings.n_sample);
     fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, 1e-3 * ctx->t_p_eval_us, n_p_eval, 1e-3 * ctx->t_p_eval_us / n_p_eval, 1e6 / ctx->t_p_eval_us * n_p_eval);
+            __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval);
     fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, 1e-3 * ctx->t_eval_us,   n_eval,   1e-3 * ctx->t_eval_us   / n_eval,   1e6 / ctx->t_eval_us   * n_eval);
-    fprintf(stderr, "%s:       total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0);
+            __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval);
+    fprintf(stderr, "%s:       total time = %8.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms));
 }
 
 void llama_reset_timings(struct llama_context * ctx) {

data/ext/llama_cpp/src/llama.h

@@ -134,6 +134,20 @@ extern "C" {
         bool quantize_output_tensor; // quantize output.weight
     } llama_model_quantize_params;
 
+    // performance timing information
+    struct llama_timings {
+        double t_start_ms;
+        double t_end_ms;
+        double t_load_ms;
+        double t_sample_ms;
+        double t_p_eval_ms;
+        double t_eval_ms;
+
+        int32_t n_sample;
+        int32_t n_p_eval;
+        int32_t n_eval;
+    };
+
     LLAMA_API struct llama_context_params llama_context_default_params();
     LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
 
@@ -331,6 +345,7 @@ extern "C" {
     LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
 
     // Performance information
+    LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
     LLAMA_API void llama_print_timings(struct llama_context * ctx);
     LLAMA_API void llama_reset_timings(struct llama_context * ctx);
 

data/lib/llama_cpp/version.rb

@@ -3,8 +3,8 @@
 # llama_cpp.rb provides Ruby bindings for the llama.cpp.
 module LLaMACpp
   # The version of llama_cpp.rb you install.
-  VERSION = '0.3.1'
+  VERSION = '0.3.2'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'master-b8c8dda'
+  LLAMA_CPP_VERSION = 'master-481f793'
 end

data/lib/llama_cpp.rb

@@ -16,8 +16,22 @@ module LLaMACpp
   # @param prompt [String] The prompt to start generation with.
   # @param n_predict [Integer] The number of tokens to predict.
   # @param n_threads [Integer] The number of threads.
+  # @param n_keep [Integer] The number of tokens to keep in the context.
+  # @param n_batch [Integer] The number of tokens to process in a batch.
+  # @param repeat_last_n [Integer] The number of tokens to consider for repetition penalty.
+  # @param repeat_penalty [Float] The repetition penalty.
+  # @param frequency [Float] The frequency penalty.
+  # @param presence [Float] The presence penalty.
+  # @param top_k [Integer] The number of tokens to consider for top-k sampling.
+  # @param top_p [Float] The probability threshold for nucleus sampling.
+  # @param tfs_z [Float] The z parameter for tail-free sampling.
+  # @param typical_p [Float] The probability for typical sampling.
+  # @param temperature [Float] The temperature for temperature sampling.
   # @return [String]
-  def generate(context, prompt, n_predict: 128, n_threads: 1) # rubocop:disable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/PerceivedComplexity
+  def generate(context, prompt, # rubocop:disable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/ParameterLists, Metrics/PerceivedComplexity
+               n_predict: 128, n_threads: 1, n_keep: 10, n_batch: 512, repeat_last_n: 64,
+               repeat_penalty: 1.1, frequency: 0.0, presence: 0.0, top_k: 40,
+               top_p: 0.95, tfs_z: 1.0, typical_p: 1.0, temperature: 0.8)
     raise ArgumentError, 'context must be an instance of LLaMACpp::Context' unless context.is_a?(LLaMACpp::Context)
     raise ArgumentError, 'prompt must be a String' unless prompt.is_a?(String)
 
@@ -31,19 +45,8 @@ module LLaMACpp
 
     embd = []
     n_consumed = 0
-    n_keep = 10
     n_past = 0
     n_remain = n_predict
-    repeat_last_n = 64
-    repeat_penalty = 1.1
-    frequency = 0.0
-    presence = 0.0
-    top_k = 40
-    top_p = 0.95
-    tfs_z = 1.0
-    typical_p = 1.0
-    temperature = 0.8
-    n_batch = 512
     n_vocab = context.n_vocab
     output = []
 

data/sig/llama_cpp.rbs

@@ -28,7 +28,10 @@ module LLaMACpp
 
   def self?.init_backend: (?numa: bool) -> void
   def self?.model_quantize: (input_path: String, output_path: String, params: ModelQuantizeParams) -> void
-  def self?.generate: (::LLaMACpp::Context, String, ?n_predict: Integer, ?n_threads: Integer) -> String
+  def self?.generate: (::LLaMACpp::Context, String,
+    ?n_predict: Integer, ?n_threads: Integer, ?n_keep: Integer, ?n_batch: Integer,
+    ?repeat_last_n: Integer, ?repeat_penalty: Float, ?frequency: Float, ?presence: Float,
+    ?top_k: Integer, ?top_p: Float, ?tfs_z: Float, ?typical_p: Float, ?temperature: Float) -> String
   def self?.print_system_info: () -> void
   def self?.token_bos: () -> Integer
   def self?.token_eos: () -> Integer
@@ -67,6 +70,20 @@ module LLaMACpp
     def apply_lora_from_file: (lora_path: String, ?base_model_path: String, ?n_threads: Integer) -> void
   end
 
+  class Timings
+    public
+
+    def t_start_ms: () -> Float
+    def t_end_ms: () -> Float
+    def t_load_ms: () -> Float
+    def t_sample_ms: () -> Float
+    def t_p_eval_ms: () -> Float
+    def t_eval_ms: () -> Float
+    def n_sample: () -> Integer
+    def n_p_eval: () -> Integer
+    def n_eval: () -> Integer
+  end
+
   class Context
     public
 
@@ -80,6 +97,7 @@ module LLaMACpp
     def n_embd: () -> Integer
     def n_vocab: () -> Integer
     def vocab: (capacity: Integer) -> [Array[String], Array[Float]]
+    def timings: () -> ::LLaMACpp::Timings
     def print_timings: () -> void
     def reset_timings: () -> void
     def token_to_str: (Integer) -> String