Merge branch 'master' into finetune-lora

README.md

@@ -729,8 +729,6 @@ python3 convert.py pygmalion-7b/ --outtype q4_1
   - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGML)
   - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML)
   - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGML)
-- Specify `-eps 1e-5` for best generation quality
-- Specify `-gqa 8` for 70B models to work
 
 ### Verifying the model files
 

examples/perplexity/perplexity.cpp

@@ -142,6 +142,14 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    if (int(tokens.size()) < 2*params.n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*params.n_ctx,
+                params.n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return {std::move(tokens), 0., {}, {}};
+    }
+
     std::vector<float>       logit_history;
     std::vector<float>       prob_history;
 
@@ -274,6 +282,13 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
     auto tim2 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
 
+    if (int(tokens.size()) < 2*params.n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*params.n_ctx,
+                params.n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return {std::move(tokens), 0., {}, {}};
+    }
+
     std::vector<float> logit_history;
     logit_history.resize(tokens.size());
 

ggml-alloc.c

@@ -321,8 +321,7 @@ bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
 //////////// compute graph allocator
 
 static bool ggml_is_view(struct ggml_tensor * t) {
-    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
+    return t->view_src != NULL;
-           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
 }
 
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
@@ -340,28 +339,6 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml
     return true;
 }
 
-static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
-    switch (t->op) {
-        case GGML_OP_PERMUTE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_VIEW:
-            return t->src[0];
-        case GGML_OP_CPY:
-            return t->src[1];
-        default:
-            return NULL;
-    }
-}
-
-static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
-    struct ggml_tensor * parent = t;
-    do {
-        parent = get_view_parent(parent);
-    } while (ggml_is_view(parent));
-    return parent;
-}
-
 static bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {
         case GGML_OP_SCALE:
@@ -369,7 +346,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_ADD:
         case GGML_OP_ADD1:
-        case GGML_OP_ACC:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -379,7 +355,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_UNARY:
         case GGML_OP_ROPE:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_SET:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_CONT:
             return true;
@@ -393,24 +368,8 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     struct hash_node * ht = alloc->hash_table;
     if (node->data == NULL) {
         if (ggml_is_view(node)) {
-            size_t offset;
+            assert(node->view_src->data != NULL);
-            switch(node->op) {
+            node->data = (char *)node->view_src->data + node->view_offs;
-                case GGML_OP_VIEW:
-                    memcpy(&offset, node->op_params, sizeof(size_t));
-                    node->data = (char *) node->src[0]->data + offset;
-                    break;
-                case GGML_OP_PERMUTE:
-                case GGML_OP_RESHAPE:
-                case GGML_OP_TRANSPOSE:
-                    node->data = node->src[0]->data;
-                    break;
-                case GGML_OP_CPY:
-                    node->data = node->src[1]->data;
-                    break;
-                default:
-                    GGML_ASSERT(!"unknown view op");
-                    break;
-            }
         } else {
             // see if we can reuse a parent's buffer (inplace)
             if (ggml_op_can_inplace(node->op)) {
@@ -430,7 +389,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     struct hash_node * p_hn = hash_get(ht, parent);
                     if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
                         if (ggml_is_view(parent)) {
-                            struct ggml_tensor * view_src = get_view_source(parent);
+                            struct ggml_tensor * view_src = parent->view_src;
                             struct hash_node * view_src_hn = hash_get(ht, view_src);
                             if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
                                 // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
@@ -472,7 +431,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
             struct ggml_tensor * node = gf->nodes[i];
 
             if (ggml_is_view(node)) {
-                struct ggml_tensor * view_src = get_view_source(node);
+                struct ggml_tensor * view_src = node->view_src;
                 hash_get(ht, view_src)->n_views += 1;
             }
 
@@ -557,7 +516,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
 
                         if (p_hn->n_children == 0 && p_hn->n_views == 0) {
                             if (ggml_is_view(parent)) {
-                                struct ggml_tensor * view_src = get_view_source(parent);
+                                struct ggml_tensor * view_src = parent->view_src;
                                 struct hash_node * view_src_hn = hash_get(ht, view_src);
                                 view_src_hn->n_views -= 1;
                                 AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);

ggml.c

@@ -4104,16 +4104,11 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) {
 }
 
 size_t ggml_nbytes(const struct ggml_tensor * tensor) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+    size_t nbytes = tensor->ne[0]*tensor->nb[0]/ggml_blck_size(tensor->type);
-
+    for (int i = 1; i < GGML_MAX_DIMS; ++i) {
-    // this should handle cases where the tensor is not contiguous in memory
+        nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
-    // probaby just:
+    }
-    //
+    return nbytes;
-    //     return tensor->ne[3]*tensor->nb[3]
-    //
-    // is enough, but just in case, adding the second part
-
-    return MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*ggml_type_size(tensor->type))/ggml_blck_size(tensor->type));
 }
 
 size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
@@ -4567,20 +4562,33 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         enum   ggml_type      type,
         int                   n_dims,
         const int64_t       * ne,
-        void                * data) {
+        struct ggml_tensor  * view_src,
+        size_t                view_offs) {
 
     assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
 
-    size_t data_size = 0;
+    // find the base tensor and absolute offset
+    if (view_src != NULL && view_src->view_src != NULL) {
+        view_offs += view_src->view_offs;
+        view_src   = view_src->view_src;
+    }
 
-    if (data == NULL && !ctx->no_alloc) {
+    size_t data_size = ggml_type_size(type)*(ne[0]/ggml_blck_size(type));
-        data_size += ggml_type_size(type)*(ne[0]/ggml_blck_size(type));
     for (int i = 1; i < n_dims; i++) {
         data_size *= ne[i];
     }
+
+    GGML_ASSERT(view_src == NULL || data_size + view_offs <= ggml_nbytes(view_src));
+
+    void * data = view_src != NULL ? view_src->data : NULL;
+    if (data != NULL) {
+        data = (char *) data + view_offs;
     }
 
-    if (ctx->scratch.data != NULL && data == NULL) {
+    size_t obj_alloc_size = 0;
+
+    if (view_src == NULL && ctx->no_alloc == false) {
+        if (ctx->scratch.data != NULL) {
             // allocate tensor data in the scratch buffer
             if (ctx->scratch.offs + data_size > ctx->scratch.size) {
                 GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n",
@@ -4592,11 +4600,13 @@ static struct ggml_tensor * ggml_new_tensor_impl(
             data = (char * const) ctx->scratch.data + ctx->scratch.offs;
 
             ctx->scratch.offs += data_size;
-
+        } else {
-        data_size = 0;
+            // allocate tensor data in the context's memory pool
+            obj_alloc_size = data_size;
+        }
     }
 
-    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + data_size);
+    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size);
 
     // TODO: for recoverable errors, we would need to free the data allocated from the scratch buffer here
 
@@ -4616,7 +4626,9 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.perf_runs    =*/ 0,
         /*.perf_cycles  =*/ 0,
         /*.perf_time_us =*/ 0,
-        /*.data         =*/ (data == NULL && !ctx->no_alloc) ? (void *)(result + 1) : data,
+        /*.view_src     =*/ view_src,
+        /*.view_offs    =*/ view_offs,
+        /*.data         =*/ obj_alloc_size > 0 ? (void *)(result + 1) : data,
         /*.name         =*/ { 0 },
         /*.extra        =*/ NULL,
         /*.padding      =*/ { 0 },
@@ -4640,28 +4652,12 @@ static struct ggml_tensor * ggml_new_tensor_impl(
     return result;
 }
 
-static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
-    GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
-    assert(params_size <= GGML_MAX_OP_PARAMS);
-    memcpy(tensor->op_params, params, params_size);
-}
-
-static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_t i) {
-    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
-    return ((const int32_t *)(tensor->op_params))[i];
-}
-
-static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
-    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
-    ((int32_t *)(tensor->op_params))[i] = value;
-}
-
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type      type,
         int                   n_dims,
         const int64_t       * ne) {
-    return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL);
+    return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL, 0);
 }
 
 struct ggml_tensor * ggml_new_tensor_1d(
@@ -4726,7 +4722,23 @@ struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value) {
 }
 
 struct ggml_tensor * ggml_dup_tensor(struct ggml_context * ctx, const struct ggml_tensor * src) {
-    return ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, NULL);
+    return ggml_new_tensor(ctx, src->type, src->n_dims, src->ne);
+}
+
+static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
+    GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
+    assert(params_size <= GGML_MAX_OP_PARAMS);
+    memcpy(tensor->op_params, params, params_size);
+}
+
+static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_t i) {
+    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
+    return ((const int32_t *)(tensor->op_params))[i];
+}
+
+static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
+    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
+    ((int32_t *)(tensor->op_params))[i] = value;
 }
 
 struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) {
@@ -5183,14 +5195,13 @@ struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char *
 
 struct ggml_tensor * ggml_view_tensor(
         struct ggml_context * ctx,
-        const struct ggml_tensor * src) {
+        struct ggml_tensor  * src) {
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src, 0);
     ggml_format_name(result, "%s (view)", src->name);
 
-    result->nb[0] = src->nb[0];
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-    result->nb[1] = src->nb[1];
+        result->nb[i] = src->nb[i];
-    result->nb[2] = src->nb[2];
+    }
-    result->nb[3] = src->nb[3];
 
     return result;
 }
@@ -5799,7 +5810,7 @@ struct ggml_tensor * ggml_repeat_back(
 
 // ggml_concat
 
-struct ggml_tensor* ggml_concat(
+struct ggml_tensor * ggml_concat(
     struct ggml_context* ctx,
     struct ggml_tensor* a,
     struct ggml_tensor* b) {
@@ -6408,7 +6419,7 @@ struct ggml_tensor * ggml_reshape(
         //GGML_ASSERT(false);
     }
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6432,7 +6443,7 @@ struct ggml_tensor * ggml_reshape_1d(
     }
 
     const int64_t ne[1] = { ne0 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6457,7 +6468,7 @@ struct ggml_tensor * ggml_reshape_2d(
     }
 
     const int64_t ne[2] = { ne0, ne1 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6483,7 +6494,7 @@ struct ggml_tensor * ggml_reshape_3d(
     }
 
     const int64_t ne[3] = { ne0, ne1, ne2 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6493,7 +6504,6 @@ struct ggml_tensor * ggml_reshape_3d(
     return result;
 }
 
-
 struct ggml_tensor * ggml_reshape_4d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -6511,7 +6521,7 @@ struct ggml_tensor * ggml_reshape_4d(
     }
 
     const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6521,34 +6531,12 @@ struct ggml_tensor * ggml_reshape_4d(
     return result;
 }
 
-// ggml_view_1d
+static struct ggml_tensor * ggml_view_impl(
-
-static struct ggml_tensor * ggml_view_tensor_offset(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         int                   n_dims,
         const int64_t       * ne,
         size_t                offset) {
-    // don't calculate an offset from an unallocated tensor
-    void * data = NULL;
-    if (a->data != NULL) {
-        data = (char *) a->data + offset;
-    }
-
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, data);
-
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
-
-    return result;
-}
-
-struct ggml_tensor * ggml_view_1d(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        int64_t               ne0,
-        size_t                offset) {
 
     bool is_node = false;
 
@@ -6556,7 +6544,10 @@ struct ggml_tensor * ggml_view_1d(
         is_node = true;
     }
 
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 1, &ne0, offset);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, a, offset);
+    ggml_format_name(result, "%s (view)", a->name);
+
+    ggml_set_op_params(result, &offset, sizeof(offset));
 
     result->op   = GGML_OP_VIEW;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6565,6 +6556,19 @@ struct ggml_tensor * ggml_view_1d(
     return result;
 }
 
+// ggml_view_1d
+
+struct ggml_tensor * ggml_view_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        size_t                offset) {
+
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 1, &ne0, offset);
+
+    return result;
+}
+
 // ggml_view_2d
 
 struct ggml_tensor * ggml_view_2d(
@@ -6575,24 +6579,14 @@ struct ggml_tensor * ggml_view_2d(
         size_t                nb1,
         size_t                offset) {
 
-    bool is_node = false;
+    const int64_t ne[2] = { ne0, ne1 };
 
-    if (a->grad) {
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 2, ne, offset);
-        is_node = true;
-    }
-
-    const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 };
-
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 2, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = result->nb[1]*ne1;
     result->nb[3] = result->nb[2];
 
-    result->op   = GGML_OP_VIEW;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-
     return result;
 }
 
@@ -6608,24 +6602,14 @@ struct ggml_tensor * ggml_view_3d(
         size_t                nb2,
         size_t                offset) {
 
-    bool is_node = false;
+    const int64_t ne[3] = { ne0, ne1, ne2 };
 
-    if (a->grad) {
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 3, ne, offset);
-        is_node = true;
-    }
-
-    const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, 1 };
-
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 3, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = nb2;
     result->nb[3] = result->nb[2]*ne2;
 
-    result->op   = GGML_OP_VIEW;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-
     return result;
 }
 
@@ -6643,24 +6627,14 @@ struct ggml_tensor * ggml_view_4d(
         size_t                nb3,
         size_t                offset) {
 
-    bool is_node = false;
+    const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
 
-    if (a->grad) {
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 4, ne, offset);
-        is_node = true;
-    }
-
-    const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, ne3 };
-
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 4, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = nb2;
     result->nb[3] = nb3;
 
-    result->op   = GGML_OP_VIEW;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-
     return result;
 }
 
@@ -6846,7 +6820,7 @@ static struct ggml_tensor * ggml_diag_mask_inf_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    int32_t params[] = { n_past, inplace ? 1 : 0 };
+    int32_t params[] = { n_past };
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_DIAG_MASK_INF;
@@ -6863,7 +6837,6 @@ struct ggml_tensor * ggml_diag_mask_inf(
     return ggml_diag_mask_inf_impl(ctx, a, n_past, false);
 }
 
-
 struct ggml_tensor * ggml_diag_mask_inf_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -6886,7 +6859,7 @@ static struct ggml_tensor * ggml_diag_mask_zero_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    int32_t params[] = { n_past, inplace ? 1 : 0 };
+    int32_t params[] = { n_past };
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_DIAG_MASK_ZERO;
@@ -12306,7 +12279,7 @@ static void ggml_compute_forward_diag_mask_f32(
     const int nth = params->nth;
 
     const int  n_past  = ((int32_t *) dst->op_params)[0];
-    const bool inplace = (bool)((int32_t *) dst->op_params)[1];
+    const bool inplace = src0->data == dst->data;
 
     GGML_ASSERT(n_past >= 0);
 

ggml.h

@@ -479,6 +479,9 @@ extern "C" {
         int64_t perf_cycles;
         int64_t perf_time_us;
 
+        struct ggml_tensor * view_src;
+        size_t               view_offs;
+
         void * data;
 
         char name[GGML_MAX_NAME];
@@ -663,7 +666,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
 
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
-    GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src);
+    GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
 
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 

llama.cpp

@@ -3211,7 +3211,7 @@ private:
 
 struct llm_bigram_bpe {
     struct comparator {
-        bool operator()(llm_bigram_bpe & l, llm_bigram_bpe & r) {
+        bool operator()(const llm_bigram_bpe & l, const llm_bigram_bpe & r) const {
             return l.rank > r.rank || (l.rank == r.rank && l.left > r.left);
         }
     };
@@ -3359,23 +3359,22 @@ private:
     }
 
     // probably not 100% correct
-    // TODO: this is quite slow - how to make it more efficient?
+    static std::vector<std::string> bpe_gpt2_preprocess(const std::string & text) {
-    static std::vector<std::string> bpe_gpt2_preprocess(std::string text) {
         std::vector<std::string> words;
 
         // ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53
         const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
         const std::regex re(pattern);
-        std::smatch m;
 
-        while (std::regex_search(text, m, re)) {
+        auto words_begin = std::sregex_iterator(text.begin(), text.end(), re);
-            for (auto x : m) {
+        auto words_end = std::sregex_iterator();
-                words.push_back(x);
+        auto n_words = std::distance(words_begin, words_end);
+        words.reserve(n_words);
+        for (auto it = words_begin; it != words_end; ++it) {
+            words.push_back(it->str());
         }
-            text = m.suffix();
-        }
-
         return words;
+
     }
 
     const llama_vocab & vocab;