llama_cpp 0.5.0 → 0.5.1

data/ext/llama_cpp/src/k_quants.c

@@ -13,6 +13,26 @@
 //
 #include <arm_neon.h>
 
+#if !defined(__aarch64__)
+inline static int32_t vaddvq_s16(int16x8_t v) {
+    return
+        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
+        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
+        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
+        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+#endif
+
 #else
 
 #ifdef __wasm_simd128__
@@ -63,7 +83,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         float ax = fabsf(x[i]);
         if (ax > amax) { amax = ax; max = x[i]; }
     }
-    if (!amax) { // all zero
+    if (amax < 1e-30f) { // all zero
         for (int i = 0; i < n; ++i) {
             L[i] = 0;
         }
@@ -183,13 +203,9 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
         int ntry, float alpha) {
     float min = x[0];
     float max = x[0];
-    float sum_x = 0;
-    float sum_x2 = 0;
     for (int i = 1; i < n; ++i) {
         if (x[i] < min) min = x[i];
         if (x[i] > max) max = x[i];
-        sum_x += x[i];
-        sum_x2 += x[i]*x[i];
     }
     if (max == min) {
         for (int i = 0; i < n; ++i) L[i] = 0;
@@ -1070,6 +1086,13 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict
 
         }
 
+        if (!max_abs_scale) {
+            memset(&y[i], 0, sizeof(block_q6_K));
+            y[i].d = ggml_fp32_to_fp16(0.f);
+            x += QK_K;
+            continue;
+        }
+
         float iscale = -128.f/max_scale;
         y[i].d = ggml_fp32_to_fp16(1/iscale);
         for (int ib = 0; ib < QK_K/16; ++ib) {
@@ -1306,7 +1329,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     const uint8x16_t m3 = vdupq_n_u8(0x3);
     const uint8x16_t m4 = vdupq_n_u8(0xF);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     int8x16x2_t q2bytes;
     uint8_t aux[16];
@@ -1612,7 +1637,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m3 = vdupq_n_u8(0x3);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q2bytes;
 
@@ -2060,7 +2087,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
     __m256 acc = _mm256_setzero_ps();
 
-    uint32_t *aux;
+    const uint32_t *aux;
 
     for (int i = 0; i < nb; ++i) {
 
@@ -2070,7 +2097,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const int8_t  * restrict q8 = y[i].qs;
 
         // Set up scales
-        aux = (uint32_t *)x[i].scales;
+        aux = (const uint32_t *)x[i].scales;
         __m128i scales128 = _mm_set_epi32(
                 ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
                 ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
@@ -2596,8 +2623,6 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         const uint8_t * restrict q4 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
 
-        //int32x4_t isum = mzero;
-
         int32_t sumi1 = 0;
         int32_t sumi2 = 0;
 
@@ -3096,9 +3121,11 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
     const uint8x16_t mone = vdupq_n_u8(1);
     const uint8x16_t mtwo = vdupq_n_u8(2);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t mzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q5bytes;
 
@@ -3441,8 +3468,10 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
     const uint8x16_t mh = vdupq_n_u8(16);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t mzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q5bytes;
     uint8x16x4_t q5h;
@@ -3660,7 +3689,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
     float sum = 0;
 
     const uint8x16_t m4b = vdupq_n_u8(0xF);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
     //const int8x16_t  m32s = vdupq_n_s8(32);
 
     const uint8x16_t mone = vdupq_n_u8(3);
@@ -4049,8 +4080,10 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
     float sum = 0;
 
     const uint8x16_t m4b = vdupq_n_u8(0xF);
-    const int32x4_t  vzero = vdupq_n_s32(0);
     const int8x16_t  m32s = vdupq_n_s8(32);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     const uint8x16_t mone = vdupq_n_u8(3);
 

data/ext/llama_cpp/src/llama.cpp

@@ -126,6 +126,9 @@ void replace_all(std::string & s, const std::string & search, const std::string
     }
     s = std::move(result);
 }
+#ifdef GGML_USE_CPU_HBM
+#include <hbwmalloc.h>
+#endif
 
 static void zeros(std::ofstream & file, size_t n) {
     char zero = 0;
@@ -325,6 +328,44 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_GPT2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+        },
+    },
+    {
+        LLM_ARCH_GPTJ,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+        },
+    },
+    {
+        LLM_ARCH_GPTNEOX,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_MPT,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+        },
+    },
+    {
+        LLM_ARCH_UNKNOWN,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+        },
+    },
 };
 
 static llm_arch llm_arch_from_string(const std::string & name) {
@@ -412,6 +453,9 @@ static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph *
 #elif GGML_USE_METAL
 #   define llama_host_malloc(n)  ggml_metal_host_malloc(n)
 #   define llama_host_free(data) ggml_metal_host_free(data)
+#elif GGML_USE_CPU_HBM
+#   define llama_host_malloc(n)  hbw_malloc(n)
+#   define llama_host_free(data) if (data != NULL) hbw_free(data)
 #else
 #   define llama_host_malloc(n)  malloc(n)
 #   define llama_host_free(data) free(data)
@@ -568,16 +612,16 @@ struct llama_mmap {
 
         if (prefetch > 0) {
             // Advise the kernel to preload the mapped memory
-            if (madvise(addr, std::min(file->size, prefetch), MADV_WILLNEED)) {
-                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+            if (posix_madvise(addr, std::min(file->size, prefetch), POSIX_MADV_WILLNEED)) {
+                fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
                         strerror(errno));
             }
         }
         if (numa) {
             // advise the kernel not to use readahead
             // (because the next page might not belong on the same node)
-            if (madvise(addr, file->size, MADV_RANDOM)) {
-                fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
+            if (posix_madvise(addr, file->size, POSIX_MADV_RANDOM)) {
+                fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
                         strerror(errno));
             }
         }
@@ -614,7 +658,9 @@ struct llama_mmap {
         #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
         if (prefetch) {
             // Advise the kernel to preload the mapped memory
+
             WIN32_MEMORY_RANGE_ENTRY range;
+
             range.VirtualAddress = addr;
             range.NumberOfBytes = (SIZE_T)size;
             if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
@@ -1446,7 +1492,11 @@ struct llama_model_loader {
             // allocate temp buffer if not using mmap
             if (!use_mmap && cur->data == NULL) {
                 GGML_ASSERT(cur->backend != GGML_BACKEND_CPU);
-                cur->data = malloc(ggml_nbytes(cur));
+                #ifdef GGML_USE_CPU_HBM
+                cur->data = (uint8_t*)hbw_malloc(ggml_nbytes(cur));
+                #else
+                cur->data = (uint8_t*)malloc(ggml_nbytes(cur));
+                #endif
             }
 
             load_data_for(cur);
@@ -1600,9 +1650,13 @@ static void llm_load_hparams(
 
         GGUF_GET_KEY(ctx, hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT));
 
-        if (hparams.n_rot != hparams.n_embd / hparams.n_head) {
-            throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head));
+        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
+            if (hparams.n_rot != hparams.n_embd / hparams.n_head) {
+                throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head));
+            }
         }
+        // gpt-neox n_rot = rotary_pct * (n_embd / n_head)
+        // gpt-j n_rot = rotary_dim
     }
 
     // arch-specific KVs
@@ -2895,7 +2949,12 @@ static bool llama_eval_internal(
 
     // for big prompts, if BLAS is enabled, it is better to use only one thread
     // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
-    n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads;
+    // TODO: this is mostly important for Apple Silicon where CBLAS is still performing very well
+    //       we still need some threads to process all non-mul_mat ops, but not too much to avoid interfering
+    //       with the BLAS calls. need a better solution
+    if (N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas()) {
+        n_threads = std::min(4, n_threads);
+    }
 
     struct ggml_tensor * res        = gf->nodes[gf->n_nodes - 1];
     struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
@@ -3000,33 +3059,10 @@ static bool llama_is_control_token(const llama_vocab & vocab, llama_token id) {
     return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_CONTROL;
 }
 
-static bool llama_is_user_defined_token(const llama_vocab & vocab, llama_token id) {
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_USER_DEFINED;
-}
-
-static bool llama_is_unused_token(const llama_vocab & vocab, llama_token id) {
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_UNUSED;
-}
-
 static bool llama_is_byte_token(const llama_vocab & vocab, llama_token id) {
     return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_BYTE;
 }
 
-static bool llama_is_bos_token(const llama_vocab & vocab, llama_token id) {
-    GGML_ASSERT(llama_is_control_token(vocab, id));
-    return id == vocab.special_bos_id;
-}
-
-static bool llama_is_eos_token(const llama_vocab & vocab, llama_token id ) {
-    GGML_ASSERT(llama_is_control_token(vocab, id));
-    return id == vocab.special_eos_id;
-}
-
-static bool llama_is_pad_token(const llama_vocab & vocab, llama_token id ) {
-    GGML_ASSERT(id < 0 || llama_is_control_token(vocab, id));
-    return id == vocab.special_pad_id;
-}
-
 static uint8_t llama_token_to_byte(const llama_vocab & vocab, llama_token id) {
     GGML_ASSERT(llama_is_byte_token(vocab, id));
     const auto& token_data = vocab.id_to_token.at(id);
@@ -3319,9 +3355,15 @@ struct llm_tokenizer_bpe {
                         std::string byte_str(1, *j);
                         auto token_multibyte = vocab.token_to_id.find(byte_str);
                         if (token_multibyte == vocab.token_to_id.end()) {
-                            fprintf(stderr,"ERROR: byte not found in vocab: '%s'\n", byte_str.c_str());
+                            try {
+                                llama_token token_byte = llama_byte_to_token(vocab, *j);
+                                output.push_back(token_byte);
+                            } catch (const std::out_of_range & err) {
+                                fprintf(stderr,"ERROR: byte not found in vocab: '%s'\n", byte_str.c_str());
+                            }
+                        } else {
+                            output.push_back((*token_multibyte).second);
                         }
-                        output.push_back((*token_multibyte).second);
                     }
                 } else {
                     output.push_back((*token).second);
@@ -3595,7 +3637,7 @@ static void llama_grammar_advance_stack(
         std::vector<std::vector<const llama_grammar_element *>> & new_stacks) {
 
     if (stack.empty()) {
-        new_stacks.push_back(stack);
+        new_stacks.emplace_back(stack);
         return;
     }
 
@@ -3632,7 +3674,7 @@ static void llama_grammar_advance_stack(
         }
         case LLAMA_GRETYPE_CHAR:
         case LLAMA_GRETYPE_CHAR_NOT:
-            new_stacks.push_back(stack);
+            new_stacks.emplace_back(stack);
             break;
         default:
             // end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range
@@ -3797,6 +3839,25 @@ void llama_grammar_free(struct llama_grammar * grammar) {
     delete grammar;
 }
 
+struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar) {
+    llama_grammar * result = new llama_grammar{ grammar->rules, grammar->stacks, grammar->partial_utf8 };
+
+    // redirect elements in stacks to point to new rules
+    for (size_t is = 0; is < result->stacks.size(); is++) {
+        for (size_t ie = 0; ie < result->stacks[is].size(); ie++) {
+            for (size_t ir0 = 0; ir0 < grammar->rules.size(); ir0++) {
+                for (size_t ir1 = 0; ir1 < grammar->rules[ir0].size(); ir1++) {
+                    if (grammar->stacks[is][ie] == &grammar->rules[ir0][ir1]) {
+                         result->stacks[is][ie]  =  &result->rules[ir0][ir1];
+                    }
+                }
+            }
+        }
+    }
+
+    return result;
+}
+
 //
 // sampling
 //
@@ -4388,7 +4449,7 @@ struct llama_logit_info {
         }
         return min_heap;
     }
-    float probability_from_logit(float logit) {
+    float probability_from_logit(float logit) const {
         return normalizer * std::exp(logit - max_l);
     }
 };
@@ -4678,6 +4739,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     llm_load_arch(*ml, model);
     llm_load_hparams(*ml, model, 0, 0, 0);
 
+    if (params->only_copy) {
+        ftype = model.ftype;
+    }
+
     const size_t align = GGUF_DEFAULT_ALIGNMENT;
     struct gguf_context * ctx_out = gguf_init_empty();
 
@@ -4719,9 +4784,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     std::vector<std::thread> workers;
     std::mutex mutex;
 
+#ifdef GGML_USE_K_QUANTS
     auto use_more_bits = [] (int i_layer, int num_layers) -> bool {
         return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2;
     };
+#endif
 
     int idx = 0;
 
@@ -4764,18 +4831,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         // quantize only 2D tensors
         quantize &= (tensor->n_dims == 2);
         quantize &= params->quantize_output_tensor || name != "output.weight";
-        quantize &= quantized_type != tensor->type;
+        quantize &= !params->only_copy;
 
         enum ggml_type new_type;
         void * new_data;
         size_t new_size;
 
-        if (!quantize) {
-            new_type = tensor->type;
-            new_data = tensor->data;
-            new_size = ggml_nbytes(tensor);
-            LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0);
-        } else {
+        if (quantize) {
             new_type = quantized_type;
 #ifdef GGML_USE_K_QUANTS
             // TODO: avoid hardcoded tensor names - use the TN_* constants
@@ -4874,7 +4936,16 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 }
             }
 #endif
-
+            // If we've decided to quantize to the same type the tensor is already
+            // in then there's nothing to do.
+            quantize = tensor->type != new_type;
+        }
+        if (!quantize) {
+            new_type = tensor->type;
+            new_data = tensor->data;
+            new_size = ggml_nbytes(tensor);
+            LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0);
+        } else {
             const size_t nelements = ggml_nelements(tensor);
 
             float * f32_data;
@@ -5279,7 +5350,7 @@ struct llama_context_params llama_context_default_params() {
         /*.seed                        =*/ LLAMA_DEFAULT_SEED,
         /*.n_ctx                       =*/ 512,
         /*.n_batch                     =*/ 512,
-        /*.gpu_layers                  =*/ 0,
+        /*.n_gpu_layers                =*/ 0,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ nullptr,
         /*.rope_freq_base              =*/ 10000.0f,
@@ -5296,6 +5367,10 @@ struct llama_context_params llama_context_default_params() {
         /*.embedding                   =*/ false,
     };
 
+#ifdef GGML_USE_METAL
+    result.n_gpu_layers = 1;
+#endif
+
     return result;
 }
 
@@ -5305,6 +5380,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
         /*.ftype                       =*/ LLAMA_FTYPE_MOSTLY_Q5_1,
         /*.allow_requantize            =*/ false,
         /*.quantize_output_tensor      =*/ true,
+        /*.only_copy                   =*/ false,
     };
 
     return result;
@@ -5487,43 +5563,43 @@ struct llama_context * llama_new_context_with_model(
             }
 #endif
         }
-    }
 
 #ifdef GGML_USE_METAL
-    if (params.n_gpu_layers > 0) {
-        // this allocates all Metal resources and memory buffers
+        if (params.n_gpu_layers > 0) {
+            // this allocates all Metal resources and memory buffers
 
-        void * data_ptr  = NULL;
-        size_t data_size = 0;
+            void * data_ptr  = NULL;
+            size_t data_size = 0;
 
-        if (params.use_mmap) {
-            data_ptr  = ctx->model.mapping->addr;
-            data_size = ctx->model.mapping->size;
-        } else {
-            data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
-            data_size = ggml_get_mem_size  (ctx->model.ctx);
-        }
+            if (params.use_mmap) {
+                data_ptr  = ctx->model.mapping->addr;
+                data_size = ctx->model.mapping->size;
+            } else {
+                data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
+                data_size = ggml_get_mem_size  (ctx->model.ctx);
+            }
 
-        const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
+            const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
 
-        LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
+            LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
 
 #define LLAMA_METAL_CHECK_BUF(result)                            \
-    if (!(result)) {                                             \
-        LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
-        llama_free(ctx);                                         \
-        return NULL;                                             \
-    }
+            if (!(result)) {                                             \
+                LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
+                llama_free(ctx);                                         \
+                return NULL;                                             \
+            }
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.data, ctx->buf_compute.size, 0));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.data, ctx->buf_compute.size, 0));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0));
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0));
 #undef LLAMA_METAL_CHECK_BUF
-    }
+        }
 #endif
+    }
 
 #ifdef GGML_USE_MPI
     ctx->ctx_mpi = ggml_mpi_init();
@@ -5857,7 +5933,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
         rng_ss.str(std::string(&rng_buf[0], rng_size));
         rng_ss >> ctx->rng;
 
-        GGML_ASSERT(rng_ss.fail() == false);
+        GGML_ASSERT(!rng_ss.fail());
     }
 
     // set logits

data/ext/llama_cpp/src/llama.h

@@ -164,6 +164,7 @@ extern "C" {
         enum llama_ftype ftype;      // quantize to this llama_ftype
         bool allow_requantize;       // allow quantizing non-f32/f16 tensors
         bool quantize_output_tensor; // quantize output.weight
+        bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
     } llama_model_quantize_params;
 
     // grammar types
@@ -409,6 +410,8 @@ extern "C" {
 
     LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
 
+    LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar);
+
     //
     // Sampling functions
     //

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 = '0.5.0'
+  VERSION = '0.5.1'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b1140'
+  LLAMA_CPP_VERSION = 'b1198'
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.5.0
+  version: 0.5.1
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2023-09-02 00:00:00.000000000 Z
+date: 2023-09-08 00:00:00.000000000 Z
 dependencies: []
 description: llama_cpp.rb provides Ruby bindings for the llama.cpp.
 email: