Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	examples/run/run.cpp
#	scripts/sync-ggml.last

common/arg.cpp

@@ -2325,5 +2325,47 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_TTS}));
 
+    add_opt(common_arg(
+        {"--embd-bge-small-en-default"},
+        string_format("use default bge-small-en-v1.5 model (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
+            params.hf_file = "bge-small-en-v1.5-q8_0.gguf";
+            params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+            params.embd_normalize = 2;
+            params.n_ctx = 512;
+            params.verbose_prompt = true;
+            params.embedding = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
+
+    add_opt(common_arg(
+        {"--embd-e5-small-en-default"},
+        string_format("use default e5-small-v2 model (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
+            params.hf_file = "e5-small-v2-q8_0.gguf";
+            params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+            params.embd_normalize = 2;
+            params.n_ctx = 512;
+            params.verbose_prompt = true;
+            params.embedding = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
+
+    add_opt(common_arg(
+        {"--embd-gte-small-default"},
+        string_format("use default gte-small model (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
+            params.hf_file = "gte-small-q8_0.gguf";
+            params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+            params.embd_normalize = 2;
+            params.n_ctx = 512;
+            params.verbose_prompt = true;
+            params.embedding = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
+
     return ctx_arg;
 }

common/llguidance.cpp

@@ -254,10 +254,10 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g
         };
     }
 
-    return new llama_sampler{
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_llg_i,
-        /* .ctx   = */ ctx,
+        /* .ctx   = */ ctx
-    };
+    );
 }
 
 #else

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -360,21 +360,15 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #endif
 
 #if defined(__loongarch_asx)
-
-typedef union {
-    int32_t i;
-    float f;
-} ft_union;
-
 /* float type data load instructions */
-static __m128 __lsx_vreplfr2vr_s(float val) {
+static __m128 __lsx_vreplfr2vr_s(const float val) {
-    ft_union fi_tmpval = {.f = val};
+    v4f32 res = {val, val, val, val};
-    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
+    return (__m128)res;
 }
 
-static __m256 __lasx_xvreplfr2vr_s(float val) {
+static __m256 __lasx_xvreplfr2vr_s(const float val) {
-    ft_union fi_tmpval = {.f = val};
+    v8f32 res = {val, val, val, val, val, val, val, val};
-    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
+    return (__m256)res;
 }
 #endif
 

ggml/src/ggml-cpu/ggml-cpu-quants.c

@@ -502,30 +502,15 @@ static __m256i lasx_shuffle_b(__m256i a, __m256i b) {
 }
 
 static __m256i lasx_extu8_16(__m128i a) {
-    __m128i zero = __lsx_vldi(0);
+    return __lasx_vext2xv_hu_bu(____m256i(a));
-    __m128i vlo = __lsx_vilvl_b(zero, a);
-    __m128i vhi = __lsx_vilvh_b(zero, a);
-    return lasx_set_q(vhi, vlo);
 }
 
 static __m256i lasx_ext8_16(__m128i a) {
-     __m128i sign = __lsx_vslti_b(a, 0);
+    return __lasx_vext2xv_h_b(____m256i(a));
-     __m128i vlo = __lsx_vilvl_b(sign, a);
-     __m128i vhi = __lsx_vilvh_b(sign, a);
-     return lasx_set_q(vhi, vlo);
 }
 
 static __m256i lasx_ext16_32(__m128i a) {
-    __m256i tmp1;
+    return __lasx_vext2xv_w_h(____m256i(a));
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 0), 0);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 1), 1);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 2), 2);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 3), 3);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 4), 4);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 5), 5);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 6), 6);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 7), 7);
-    return tmp1;
 }
 
 static __m128i lasx_extracti128( __m256i a, int pos) {
@@ -593,12 +578,10 @@ static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
 // horizontally add 8 floats
 static inline float hsum_float_8(const __m256 x) {
     __m128 res = lasx_extractf128(x, 1);
-    ft_union tmp;
     res = __lsx_vfadd_s(res, lasx_extractf128(x, 0));
     res = __lsx_vfadd_s(res, (__m128)__lsx_vpickod_d((__m128i)res, (__m128i)res));
     res = __lsx_vfadd_s(res, (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w(res, 1), 0));
-    tmp.i = __lsx_vpickve2gr_w(res, 0);
+    return ((v4f32)res)[0];
-    return tmp.f;
 }
 
 // horizontally add 8 int32_t
@@ -940,7 +923,6 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
 
 #elif defined(__loongarch_asx)
     for (int i = 0; i < nb; i++) {
-        ft_union fi;
         __m256 v0 = (__m256)__lasx_xvld( x , 0);
         __m256 v1 = (__m256)__lasx_xvld( x , 32);
         __m256 v2 = (__m256)__lasx_xvld( x , 64);
@@ -958,8 +940,7 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
         __m128 tmp = max4;
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vinsgr2vr_w(tmp, __lsx_vpickve2gr_w( max4, 1 ), 0 ));
-        fi.i = __lsx_vpickve2gr_w( (__m128i)max4, 0 );
+        const float max_scalar = ((v4f32)max4)[0];
-        const float max_scalar = fi.f;
 
         // Quantize these floats
         const float d = max_scalar / 127.f;
@@ -1264,7 +1245,6 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
 
 #elif defined(__loongarch_asx)
     for (int i = 0; i < nb; i++) {
-        ft_union ft;
         __m256 v0 = (__m256)__lasx_xvld( x , 0 );
         __m256 v1 = (__m256)__lasx_xvld( x , 32 );
         __m256 v2 = (__m256)__lasx_xvld( x , 64 );
@@ -1282,8 +1262,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
         __m128 tmp = max4;
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
-        ft.i = __lsx_vpickve2gr_w( (__m128i)max4, 0 );
+        const float max_scalar = ((v4f32)max4)[0];
-        const float max_scalar = ft.f;
 
         // Quantize these floats
         const float d = max_scalar / 127.f;
@@ -6155,9 +6134,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     acc_m = __lsx_vfadd_s(acc_m, (__m128)tmp1);
 
 
-    ft_union fi;
+    *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
-    fi.i = __lsx_vpickve2gr_w(acc_m, 0);
-    *s = hsum_float_8(acc) + fi.f ;
 #else
 
     const uint8_t * scales = (const uint8_t*)&utmp[0];

ggml/src/ggml-cpu/ggml-cpu.c

@@ -1082,29 +1082,23 @@ do {                                                              \
 #define GGML_F16_STEP 32
 #define GGML_F16_EPR  8
 
-// F16 arithmetic is not supported by AVX, so we use F32 instead
+// F16 arithmetic is not supported by LASX, so we use F32 instead
 
 #define GGML_F32Cx8          __m256
 #define GGML_F32Cx8_ZERO    (__m256)__lasx_xvldi(0)
 #define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x))
 
 static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) {
-    float tmp[8];
+    __m256i a;
-
+    memcpy(&a, x, sizeof(ggml_fp16_t) * 8);
-    for (int i = 0; i < 8; i++) {
+    a = __lasx_xvpermi_d(a, 0 | (1 << 4));
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+    return __lasx_xvfcvtl_s_h(a);
 }
 
-    return (__m256)__lasx_xvld(tmp, 0);
-}
 static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
-    float arr[8];
+    __m256i a = __lasx_xvfcvt_h_s(y, y);
-
+    a = __lasx_xvpermi_d(a, 0 | (2 << 2));
-    __lasx_xvst(y, arr, 0);
+    memcpy(x, &a, sizeof(ggml_fp16_t) * 8);
-
-    for (int i = 0; i < 8; i++) {
-        x[i] = GGML_FP32_TO_FP16(arr[i]);
-    }
 }
 #define GGML_F32Cx8_LOAD(x)     __lasx_f32cx8_load(x)
 #define GGML_F32Cx8_STORE(x, y) __lasx_f32cx8_store(x, y)

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -1045,7 +1045,28 @@ bool rpc_server::copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_co
         ggml_free(ctx);
         return false;
     }
-    GGML_PRINT_DEBUG("[%s] src->buffer: %p, dst->buffer: %p\n", __func__, (void*)src->buffer, (void*)dst->buffer);
+
+    uint64_t src_size   = (uint64_t) ggml_nbytes(src);
+    uint64_t dst_data   = (uint64_t) dst->data;
+    uint64_t dst_base   = (uint64_t) ggml_backend_buffer_get_base(dst->buffer);
+    uint64_t dst_buf_sz = (uint64_t) ggml_backend_buffer_get_size(dst->buffer);
+
+    if (dst_data + src_size > dst_base + dst_buf_sz) {
+        GGML_PRINT_DEBUG("[%s] out-of-bounds write in rpc_server::copy_tensor:\n"
+                         "    write range : [0x%" PRIx64 ", 0x%" PRIx64 "]\n"
+                         "    buffer base: [0x%" PRIx64 ", 0x%" PRIx64 "]\n",
+                         __func__,
+                         dst_data,
+                         dst_data + src_size,
+                         dst_base,
+                         dst_base + dst_buf_sz);
+        ggml_free(ctx);
+        return false;
+    }
+
+    GGML_PRINT_DEBUG("[%s] src->buffer: %p, dst->buffer: %p\n",
+                     __func__, (void*) src->buffer, (void*) dst->buffer);
+
     response.result = ggml_backend_buffer_copy_tensor(src, dst);
     ggml_free(ctx);
     return true;

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -103,11 +103,10 @@ void print_device_detail(int id, sycl::device &device, std::string device_type)
     name = std::regex_replace(name, std::regex("\\(TM\\)"), "");
 
     auto global_mem_size = prop.get_global_mem_size()/1000000;
-    std::string xmx = gpu_has_xmx(device) ? "yes" : "no";
+    GGML_LOG_INFO("|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|\n", id, device_type.c_str(),
-    GGML_LOG_INFO("|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|%14s|\n", id, device_type.c_str(),
             name.c_str(), version.c_str(), prop.get_max_compute_units(),
             prop.get_max_work_group_size(), prop.get_max_sub_group_size(),
-            global_mem_size, device.get_info<sycl::info::device::driver_version>().c_str(), xmx.c_str());
+            global_mem_size, device.get_info<sycl::info::device::driver_version>().c_str());
 }
 
 void ggml_backend_sycl_print_sycl_devices() {
@@ -118,16 +117,16 @@ void ggml_backend_sycl_print_sycl_devices() {
 
     GGML_LOG_INFO(
         "|  |                   |                                       |      "
-        " |Max    |        |Max  |Global |                     |         XMX  |\n");
+        " |Max    |        |Max  |Global |                     |\n");
     GGML_LOG_INFO(
         "|  |                   |                                       |      "
-        " |compute|Max work|sub  |mem    |                     |          or  |\n");
+        " |compute|Max work|sub  |mem    |                     |\n");
     GGML_LOG_INFO(
         "|ID|        Device Type|                                   "
-        "Name|Version|units  |group   |group|size   |       Driver version| Tensor Cores |\n");
+        "Name|Version|units  |group   |group|size   |       Driver version|\n");
     GGML_LOG_INFO(
         "|--|-------------------|---------------------------------------|------"
-        "-|-------|--------|-----|-------|---------------------|--------------|\n");
+        "-|-------|--------|-----|-------|---------------------|\n");
 
     for (int id = 0; id < device_count; ++id) {
       sycl::device device = dpct::dev_mgr::instance().get_device(id);

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -2788,8 +2788,9 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     std::string matrix_cores = coopmat2_support ? "NV_coopmat2" : coopmat_support ? "KHR_coopmat" : "none";
 
     std::string device_name = props2.properties.deviceName.data();
-    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | matrix cores: %s\n",
+    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | shared memory: %d | matrix cores: %s\n",
-              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size, matrix_cores.c_str());
+              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size,
+              props2.properties.limits.maxComputeSharedMemorySize, matrix_cores.c_str());
 
     if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
         GGML_LOG_DEBUG("ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want.\n");

include/llama.h

@@ -1116,11 +1116,12 @@ extern "C" {
     };
 
     struct llama_sampler {
-        struct llama_sampler_i  * iface;
+        const struct llama_sampler_i * iface;
         llama_sampler_context_t        ctx;
     };
 
     // mirror of llama_sampler_i:
+    LLAMA_API struct llama_sampler * llama_sampler_init  (const struct llama_sampler_i * iface, llama_sampler_context_t ctx);
     LLAMA_API const char *           llama_sampler_name  (const struct llama_sampler * smpl);
     LLAMA_API void                   llama_sampler_accept(      struct llama_sampler * smpl, llama_token token);
     LLAMA_API void                   llama_sampler_apply (      struct llama_sampler * smpl, llama_token_data_array * cur_p);

src/llama-sampling.cpp

@@ -316,6 +316,13 @@ static uint32_t get_rng_seed(uint32_t seed) {
 
 // llama_sampler API
 
+struct llama_sampler * llama_sampler_init(const struct llama_sampler_i * iface, llama_sampler_context_t ctx) {
+    return new llama_sampler {
+        /* .iface = */ iface,
+        /* .ctx   = */ ctx,
+    };
+}
+
 const char * llama_sampler_name(const struct llama_sampler * smpl) {
     if (!smpl->iface) {
         return "(null)";
@@ -347,10 +354,10 @@ struct llama_sampler * llama_sampler_clone(const struct llama_sampler * smpl) {
     }
 
     if (smpl->ctx == nullptr) {
-        return new llama_sampler {
+        return llama_sampler_init(
             /* .iface = */ smpl->iface,
-            /* .ctx   = */ nullptr,
+            /* .ctx   = */ nullptr
-        };
+        );
     }
 
     GGML_ABORT("the sampler does not support cloning");
@@ -472,15 +479,15 @@ static struct llama_sampler_i llama_sampler_chain_i = {
 };
 
 struct llama_sampler * llama_sampler_chain_init(struct llama_sampler_chain_params params) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_chain_i,
         /* .ctx   = */ new llama_sampler_chain {
             /* .params      = */ params,
             /* .samplers    = */ {},
             /* .t_sample_us = */ 0,
             /* .n_sample    = */ 0,
-        },
+        }
-    };
+    );
 }
 
 void llama_sampler_chain_add(struct llama_sampler * chain, struct llama_sampler * smpl) {
@@ -546,10 +553,10 @@ static struct llama_sampler_i llama_sampler_greedy_i = {
 };
 
 struct llama_sampler * llama_sampler_init_greedy() {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_greedy_i,
-        /* .ctx   = */ nullptr,
+        /* .ctx   = */ nullptr
-    };
+    );
 }
 
 // dist
@@ -608,14 +615,14 @@ static struct llama_sampler_i llama_sampler_dist_i = {
 
 struct llama_sampler * llama_sampler_init_dist(uint32_t seed) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_dist_i,
         /* .ctx   = */ new llama_sampler_dist {
             /* .seed     = */ seed,
             /* .seed_cur = */ seed_cur,
             /* .rng      = */ std::mt19937(seed_cur),
-        },
+        }
-    };
+    );
 }
 
 // softmax
@@ -638,10 +645,10 @@ static struct llama_sampler_i llama_sampler_softmax_i = {
 };
 
 struct llama_sampler * llama_sampler_init_softmax() {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_softmax_i,
-        /* .ctx   = */ nullptr,
+        /* .ctx   = */ nullptr
-    };
+    );
 }
 
 // top-k
@@ -678,12 +685,12 @@ static struct llama_sampler_i llama_sampler_top_k_i = {
 };
 
 struct llama_sampler * llama_sampler_init_top_k(int32_t k) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_top_k_i,
         /* .ctx   = */ new llama_sampler_top_k {
             /* .k = */ k,
-        },
+        }
-    };
+    );
 }
 
 // top-p
@@ -744,13 +751,13 @@ static struct llama_sampler_i llama_sampler_top_p_i = {
 };
 
 struct llama_sampler * llama_sampler_init_top_p(float p, size_t min_keep) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_top_p_i,
         /* .ctx   = */ new llama_sampler_top_p {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
-        },
+        }
-    };
+    );
 }
 
 // min-p
@@ -840,13 +847,13 @@ static struct llama_sampler_i llama_sampler_min_p_i = {
 };
 
 struct llama_sampler * llama_sampler_init_min_p(float p, size_t min_keep) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_min_p_i,
         /* .ctx   = */ new llama_sampler_min_p {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
-        },
+        }
-    };
+    );
 }
 
 // typical
@@ -939,13 +946,13 @@ static struct llama_sampler_i llama_sampler_typical_i = {
 };
 
 struct llama_sampler * llama_sampler_init_typical(float p, size_t min_keep) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_typical_i,
         /* .ctx   = */ new llama_sampler_typical {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
-        },
+        }
-    };
+    );
 }
 
 // temp
@@ -983,12 +990,12 @@ static struct llama_sampler_i llama_sampler_temp_i = {
 };
 
 struct llama_sampler * llama_sampler_init_temp(float temp) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_temp_i,
         /* .ctx   = */ new llama_sampler_temp {
             /*.temp = */ temp,
-        },
+        }
-    };
+    );
 }
 
 // temp-ext
@@ -1093,14 +1100,14 @@ static struct llama_sampler_i llama_sampler_temp_ext_i = {
 };
 
 struct llama_sampler * llama_sampler_init_temp_ext(float temp, float delta, float exponent) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_temp_ext_i,
         /* .ctx   = */ new llama_sampler_temp_ext {
             /* .temp     = */ temp,
             /* .delta    = */ delta,
             /* .exponent = */ exponent,
-        },
+        }
-    };
+    );
 }
 
 // xtc
@@ -1185,7 +1192,7 @@ static struct llama_sampler_i llama_sampler_xtc_i = {
 
 struct llama_sampler * llama_sampler_init_xtc(float p, float t, size_t min_keep, uint32_t seed) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_xtc_i,
         /* .ctx   = */ new llama_sampler_xtc {
             /* .probability   = */ p,
@@ -1194,8 +1201,8 @@ struct llama_sampler * llama_sampler_init_xtc(float p, float t, size_t min_keep,
             /* .seed          = */ seed,
             /* .seed_cur      = */ seed_cur,
             /* .rng           = */ std::mt19937(seed_cur),
-        },
+        }
-    };
+    );
 }
 
 // mirostat
@@ -1292,7 +1299,7 @@ static struct llama_sampler_i llama_sampler_mirostat_i = {
 
 struct llama_sampler * llama_sampler_init_mirostat(int32_t n_vocab, uint32_t seed, float tau, float eta, int32_t m) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_mirostat_i,
         /* .ctx   = */ new llama_sampler_mirostat {
             /* .n_vocab  = */ n_vocab,
@@ -1303,8 +1310,8 @@ struct llama_sampler * llama_sampler_init_mirostat(int32_t n_vocab, uint32_t see
             /* .m        = */ m,
             /* .mu       = */ 2.0f*tau,
             /* .rng      = */ std::mt19937(seed_cur),
-        },
+        }
-    };
+    );
 }
 
 // mirostat v2
@@ -1391,7 +1398,7 @@ static struct llama_sampler_i llama_sampler_mirostat_v2_i = {
 
 struct llama_sampler * llama_sampler_init_mirostat_v2(uint32_t seed, float tau, float eta) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_mirostat_v2_i,
         /* .ctx   = */ new llama_sampler_mirostat_v2 {
             /* .seed     = */ seed,
@@ -1400,8 +1407,8 @@ struct llama_sampler * llama_sampler_init_mirostat_v2(uint32_t seed, float tau,
             /* .eta      = */ eta,
             /* .mu       = */ 2.0f*tau,
             /* .rng      = */ std::mt19937(seed_cur),
-        },
+        }
-    };
+    );
 }
 
 // grammar
@@ -1528,10 +1535,10 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
         };
     }
 
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_grammar_i,
-        /* .ctx   = */ ctx,
+        /* .ctx   = */ ctx
-    };
+    );
 }
 
 struct llama_sampler * llama_sampler_init_grammar(
@@ -1678,7 +1685,7 @@ struct llama_sampler * llama_sampler_init_penalties(
         float penalty_present) {
     penalty_last_n = std::max(penalty_last_n, 0);
 
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_penalties_i,
         /* .ctx   = */ new llama_sampler_penalties {
             /* .penalty_last_n  = */ penalty_last_n,
@@ -1687,8 +1694,8 @@ struct llama_sampler * llama_sampler_init_penalties(
             /* .penalty_present = */ penalty_present,
             /* .prev            = */ ring_buffer<llama_token>(penalty_last_n),
             /* .token_count     = */ {},
-        },
+        }
-    };
+    );
 }
 
 // DRY
@@ -2041,7 +2048,7 @@ struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab,
         }
     }
 
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_dry_i,
         /* .ctx   = */ new llama_sampler_dry {
             /* .total_context_size     = */ context_size,
@@ -2053,8 +2060,8 @@ struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab,
             /* .dry_repeat_count       = */ dry_enabled ? std::vector<int>(effective_dry_penalty_last_n, 0) : std::vector<int>{},
             /* .dry_max_token_repeat   = */ {},
             /* .last_tokens            = */ dry_enabled ? ring_buffer<llama_token>(effective_dry_penalty_last_n) : ring_buffer<llama_token>(0),
-        },
+        }
-    };
+    );
 }
 
 // wrapper for test-sampling.cpp
@@ -2155,14 +2162,14 @@ struct llama_sampler * llama_sampler_init_logit_bias(
                          int32_t   n_vocab,
                          int32_t   n_logit_bias,
           const llama_logit_bias * logit_bias) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_logit_bias_i,
         /* .ctx   = */ new llama_sampler_logit_bias {
             /* .n_vocab    = */ n_vocab,
             /* .logit_bias = */ std::vector<llama_logit_bias>(logit_bias, logit_bias + n_logit_bias),
             /* .to_search  = */ {},
-        },
+        }
-    };
+    );
 }
 
 // infill
@@ -2377,14 +2384,14 @@ static struct llama_sampler_i llama_sampler_infill_i = {
 };
 
 struct llama_sampler * llama_sampler_init_infill(const struct llama_vocab * vocab) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_infill_i,
         /* .ctx   = */ new llama_sampler_infill {
             /* .vocab = */ vocab,
             /* .buf0  = */ std::vector<char>(512),
             /* .buf1  = */ std::vector<char>(512),
-        },
+        }
-    };
+    );
 }
 
 // utils

src/llama.cpp

@@ -7253,7 +7253,7 @@ struct llm_build_context {
                 struct ggml_tensor * Qcur = nullptr;
                 struct ggml_tensor * Kcur = nullptr;
                 struct ggml_tensor * Vcur = nullptr;
-                if (model.type == LLM_TYPE_1_5B || model.type == LLM_TYPE_4B || model.type == LLM_TYPE_9B) {
+                if (model.layers[il].wqkv == nullptr) {
                     Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
                     if (model.layers[il].bq) {
                         Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
@@ -8837,12 +8837,14 @@ static int llama_decode_impl(
     //llama_synchronize(&lctx);
 
     // decide if we need to defrag the kv cache
-    if (cparams.causal_attn && cparams.defrag_thold >= 0.0f) {
+    if (cparams.causal_attn && cparams.defrag_thold > 0.0f) {
-        const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used)/float(kv_self.n) : 0.0f;
+        // - do not defrag small contexts (i.e. < 2048 tokens)
+        // - count the padding towards the number of used tokens
+        const float fragmentation = kv_self.n >= 2048 ? std::max(0.0f, 1.0f - float(kv_self.used + llama_kv_cache_get_padding(cparams))/float(kv_self.n)) : 0.0f;
 
         // queue defragmentation for next llama_kv_cache_update
         if (fragmentation > cparams.defrag_thold) {
-            //LLAMA_LOG_INFO("fragmentation: %.2f\n", fragmentation);
+            LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);
 
             llama_kv_cache_defrag(kv_self);
         }
@@ -9469,7 +9471,6 @@ static struct llama_model * llama_model_load_from_file_impl(
         struct llama_model_params params) {
     ggml_time_init();
 
-    llama_model * model = new llama_model(params);
 
     unsigned cur_percentage = 0;
     if (params.progress_callback == NULL) {
@@ -9488,6 +9489,8 @@ static struct llama_model * llama_model_load_from_file_impl(
         };
     }
 
+    llama_model * model = new llama_model(params);
+
     // create list of devices to use with this model
     if (params.devices) {
         for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {

src/unicode.cpp

@@ -626,8 +626,15 @@ std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
     result.reserve(utf8.size());
     size_t offset = 0;
     while (offset < utf8.size()) {
+        try {
             result.push_back(unicode_cpt_from_utf8(utf8, offset));
         }
+        catch (const std::invalid_argument & /*ex*/) {
+            // Silently ignore invalid UTF-8 input to avoid leaking the exception beyond llama_tokenize
+            ++offset;
+            result.emplace_back(0xFFFD); // replacement character
+        }
+    }
     return result;
 }