cui-llama.rn 1.3.4 → 1.3.5

package/cpp/ggml.c

@@ -3773,13 +3773,84 @@ struct lm_ggml_tensor * lm_ggml_clamp(
     return result;
 }
 
-// lm_ggml_conv_1d
-
 static int64_t lm_ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
     return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
 }
 
-LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_1d(
+// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
+// a: [OC，IC, KH, KW]
+// b: [N, IC, IH, IW]
+// result: [N, OH, OW, IC*KH*KW]
+struct lm_ggml_tensor * lm_ggml_im2col(
+        struct lm_ggml_context * ctx,
+        struct lm_ggml_tensor  * a,
+        struct lm_ggml_tensor  * b,
+        int                   s0,
+        int                   s1,
+        int                   p0,
+        int                   p1,
+        int                   d0,
+        int                   d1,
+        bool                  is_2D,
+        enum lm_ggml_type        dst_type) {
+    if (is_2D) {
+        LM_GGML_ASSERT(a->ne[2] == b->ne[2]);
+    } else {
+        //LM_GGML_ASSERT(b->ne[1] % a->ne[1] == 0);
+        LM_GGML_ASSERT(b->ne[1] == a->ne[1]);
+        LM_GGML_ASSERT(b->ne[3] == 1);
+    }
+
+    const int64_t OH = is_2D ? lm_ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0;
+    const int64_t OW =         lm_ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+
+    LM_GGML_ASSERT((!is_2D || OH > 0) && "b too small compared to a");
+    LM_GGML_ASSERT((OW > 0)           && "b too small compared to a");
+
+    const int64_t ne[4] = {
+        is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0],
+        OW,
+        is_2D ? OH : b->ne[2],
+        is_2D ?      b->ne[3] : 1,
+    };
+
+    struct lm_ggml_tensor * result = lm_ggml_new_tensor(ctx, dst_type, 4, ne);
+    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
+    lm_ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = LM_GGML_OP_IM2COL;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+struct lm_ggml_tensor * lm_ggml_im2col_back(
+        struct lm_ggml_context * ctx,
+        struct lm_ggml_tensor  * a,
+        struct lm_ggml_tensor  * b,
+        int64_t             * ne,
+        int                   s0,
+        int                   s1,
+        int                   p0,
+        int                   p1,
+        int                   d0,
+        int                   d1,
+        bool                  is_2D) {
+    struct lm_ggml_tensor * result = lm_ggml_new_tensor(ctx, LM_GGML_TYPE_F32, 4, ne);
+    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
+    lm_ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = LM_GGML_OP_IM2COL_BACK;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+// lm_ggml_conv_1d
+
+struct lm_ggml_tensor * lm_ggml_conv_1d(
         struct lm_ggml_context * ctx,
         struct lm_ggml_tensor  * a,
         struct lm_ggml_tensor  * b,
@@ -3809,137 +3880,75 @@ struct lm_ggml_tensor* lm_ggml_conv_1d_ph(
     return lm_ggml_conv_1d(ctx, a, b, s, a->ne[0] / 2, d);
 }
 
-// lm_ggml_conv_transpose_1d
+// lm_ggml_conv_1d_dw
 
-static int64_t lm_ggml_calc_conv_transpose_1d_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
-    return (ins - 1) * s - 2 * p + d * (ks - 1) + 1;
-}
-
-LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_transpose_1d(
+struct lm_ggml_tensor * lm_ggml_conv_1d_dw(
         struct lm_ggml_context * ctx,
         struct lm_ggml_tensor  * a,
         struct lm_ggml_tensor  * b,
         int                   s0,
         int                   p0,
         int                   d0) {
-    LM_GGML_ASSERT(lm_ggml_is_matrix(b));
-    LM_GGML_ASSERT(a->ne[2] == b->ne[1]);
-    LM_GGML_ASSERT(a->ne[3] == 1);
+    struct lm_ggml_tensor * new_a = lm_ggml_reshape_4d(ctx, a, a->ne[0], 1, a->ne[1], a->ne[2]);
+    struct lm_ggml_tensor * new_b = lm_ggml_reshape_4d(ctx, b, b->ne[0], 1, b->ne[1], b->ne[2]);
 
-    LM_GGML_ASSERT(p0 == 0);
-    LM_GGML_ASSERT(d0 == 1);
+    struct lm_ggml_tensor * im2col = lm_ggml_im2col(ctx, new_a, new_b, s0, 0, p0, 0, d0, 0, false, LM_GGML_TYPE_F16);
 
-    const int64_t ne[4] = {
-        lm_ggml_calc_conv_transpose_1d_output_size(b->ne[0], a->ne[0], s0, 0 /*p0*/, 1 /*d0*/),
-        a->ne[1], b->ne[2], 1,
-    };
-    struct lm_ggml_tensor * result = lm_ggml_new_tensor(ctx, LM_GGML_TYPE_F32, 4, ne);
+    struct lm_ggml_tensor * result = lm_ggml_mul_mat(ctx, im2col, a);
 
-    int32_t params[] = { s0, p0, d0 };
-    lm_ggml_set_op_params(result, params, sizeof(params));
-
-    result->op     = LM_GGML_OP_CONV_TRANSPOSE_1D;
-    result->src[0] = a;
-    result->src[1] = b;
+    result = lm_ggml_reshape_3d(ctx, result, b->ne[0], b->ne[1], 1);
 
     return result;
 }
 
-// lm_ggml_conv_depthwise
+// lm_ggml_conv_1d_dw_ph
 
-struct lm_ggml_tensor * lm_ggml_conv_depthwise_2d(
+struct lm_ggml_tensor * lm_ggml_conv_1d_dw_ph(
         struct lm_ggml_context * ctx,
         struct lm_ggml_tensor  * a,
         struct lm_ggml_tensor  * b,
         int                   s0,
-        int                   s1,
-        int                   p0,
-        int                   p1,
-        int                   d0,
-        int                   d1) {
-    struct lm_ggml_tensor * new_a = lm_ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]);
-    struct lm_ggml_tensor * im2col = lm_ggml_im2col(ctx, new_a,
-                                        lm_ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]),
-                                        s0, s1, p0, p1, d0, d1, true, LM_GGML_TYPE_F16); // [N * IC, OH, OW, KH * KW]
-    struct lm_ggml_tensor * new_b = lm_ggml_reshape_4d(ctx, im2col, im2col->ne[0], im2col->ne[2] * im2col->ne[1], b->ne[2], b->ne[3]); // [N * IC, OH, OW, KH * KW] => [N, IC, OH * OW, KH * KW]
+        int                   d0) {
+    return lm_ggml_conv_1d_dw(ctx, a, b, s0, a->ne[0] / 2, d0);
+}
 
-    new_a = lm_ggml_reshape_4d(ctx, new_a, (new_a->ne[0] * new_a->ne[1]), new_a->ne[2],  new_a->ne[3], 1);                       // [OC，1, KH, KW] => [1, OC, 1, KH * KW]
-    struct lm_ggml_tensor * result = lm_ggml_mul_mat(ctx, new_a, new_b);
-    result = lm_ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], b->ne[2], b->ne[3]); // [N, OC, OH, OW]
+// lm_ggml_conv_transpose_1d
 
-    return result;
+static int64_t lm_ggml_calc_conv_transpose_1d_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
+    return (ins - 1) * s - 2 * p + d * (ks - 1) + 1;
 }
-// lm_ggml_conv_2d
 
-// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
-// a: [OC，IC, KH, KW]
-// b: [N, IC, IH, IW]
-// result: [N, OH, OW, IC*KH*KW]
-struct lm_ggml_tensor * lm_ggml_im2col(
+LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_transpose_1d(
         struct lm_ggml_context * ctx,
         struct lm_ggml_tensor  * a,
         struct lm_ggml_tensor  * b,
         int                   s0,
-        int                   s1,
         int                   p0,
-        int                   p1,
-        int                   d0,
-        int                   d1,
-        bool                  is_2D,
-        enum lm_ggml_type        dst_type) {
-    if(is_2D) {
-        LM_GGML_ASSERT(a->ne[2] == b->ne[2]);
-    } else {
-        LM_GGML_ASSERT(a->ne[1] == b->ne[1]);
-        LM_GGML_ASSERT(b->ne[3] == 1);
-    }
-
-    const int64_t OH = is_2D ? lm_ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0;
-    const int64_t OW =         lm_ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+        int                   d0) {
+    LM_GGML_ASSERT(lm_ggml_is_matrix(b));
+    LM_GGML_ASSERT(a->ne[2] == b->ne[1]);
+    LM_GGML_ASSERT(a->ne[3] == 1);
 
-    LM_GGML_ASSERT((!is_2D || OH > 0) && "b too small compared to a");
-    LM_GGML_ASSERT((OW > 0)           && "b too small compared to a");
+    LM_GGML_ASSERT(p0 == 0);
+    LM_GGML_ASSERT(d0 == 1);
 
     const int64_t ne[4] = {
-        is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0],
-        OW,
-        is_2D ? OH : b->ne[2],
-        is_2D ?      b->ne[3] : 1,
+        lm_ggml_calc_conv_transpose_1d_output_size(b->ne[0], a->ne[0], s0, 0 /*p0*/, 1 /*d0*/),
+        a->ne[1], b->ne[2], 1,
     };
+    struct lm_ggml_tensor * result = lm_ggml_new_tensor(ctx, LM_GGML_TYPE_F32, 4, ne);
 
-    struct lm_ggml_tensor * result = lm_ggml_new_tensor(ctx, dst_type, 4, ne);
-    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
+    int32_t params[] = { s0, p0, d0 };
     lm_ggml_set_op_params(result, params, sizeof(params));
 
-    result->op     = LM_GGML_OP_IM2COL;
+    result->op     = LM_GGML_OP_CONV_TRANSPOSE_1D;
     result->src[0] = a;
     result->src[1] = b;
 
     return result;
 }
 
-struct lm_ggml_tensor * lm_ggml_im2col_back(
-        struct lm_ggml_context * ctx,
-        struct lm_ggml_tensor  * a,
-        struct lm_ggml_tensor  * b,
-        int64_t             * ne,
-        int                   s0,
-        int                   s1,
-        int                   p0,
-        int                   p1,
-        int                   d0,
-        int                   d1,
-        bool                  is_2D) {
-    struct lm_ggml_tensor * result = lm_ggml_new_tensor(ctx, LM_GGML_TYPE_F32, 4, ne);
-    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
-    lm_ggml_set_op_params(result, params, sizeof(params));
-
-    result->op     = LM_GGML_OP_IM2COL_BACK;
-    result->src[0] = a;
-    result->src[1] = b;
-
-    return result;
-}
+// lm_ggml_conv_2d
 
 // a: [OC，IC, KH, KW]
 // b: [N, IC, IH, IW]
@@ -3986,6 +3995,31 @@ struct lm_ggml_tensor * lm_ggml_conv_2d_s1_ph(
     return lm_ggml_conv_2d(ctx, a, b, 1, 1, a->ne[0] / 2, a->ne[1] / 2, 1, 1);
 }
 
+// lm_ggml_conv_2d_dw
+
+struct lm_ggml_tensor * lm_ggml_conv_2d_dw(
+        struct lm_ggml_context * ctx,
+        struct lm_ggml_tensor  * a,
+        struct lm_ggml_tensor  * b,
+        int                   s0,
+        int                   s1,
+        int                   p0,
+        int                   p1,
+        int                   d0,
+        int                   d1) {
+    struct lm_ggml_tensor * new_a = lm_ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]);
+    struct lm_ggml_tensor * im2col = lm_ggml_im2col(ctx, new_a,
+                                        lm_ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]),
+                                        s0, s1, p0, p1, d0, d1, true, LM_GGML_TYPE_F16); // [N * IC, OH, OW, KH * KW]
+    struct lm_ggml_tensor * new_b = lm_ggml_reshape_4d(ctx, im2col, im2col->ne[0], im2col->ne[2] * im2col->ne[1], b->ne[2], b->ne[3]); // [N * IC, OH, OW, KH * KW] => [N, IC, OH * OW, KH * KW]
+
+    new_a = lm_ggml_reshape_4d(ctx, new_a, (new_a->ne[0] * new_a->ne[1]), new_a->ne[2],  new_a->ne[3], 1);                       // [OC，1, KH, KW] => [1, OC, 1, KH * KW]
+    struct lm_ggml_tensor * result = lm_ggml_mul_mat(ctx, new_a, new_b);
+    result = lm_ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], b->ne[2], b->ne[3]); // [N, OC, OH, OW]
+
+    return result;
+}
+
 // lm_ggml_conv_transpose_2d_p0
 
 static int64_t lm_ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int s, int p) {
@@ -6050,12 +6084,12 @@ struct lm_ggml_tensor * lm_ggml_graph_get_tensor(const struct lm_ggml_cgraph * c
 
 struct lm_ggml_tensor * lm_ggml_graph_get_grad(const struct lm_ggml_cgraph * cgraph, const struct lm_ggml_tensor * node) {
     const size_t igrad = lm_ggml_hash_find(&cgraph->visited_hash_set, node);
-    return igrad != LM_GGML_HASHSET_FULL && lm_ggml_bitset_get(cgraph->visited_hash_set.used, igrad) ? cgraph->grads[igrad] : NULL;
+    return igrad != LM_GGML_HASHSET_FULL && lm_ggml_bitset_get(cgraph->visited_hash_set.used, igrad) && cgraph->grads ? cgraph->grads[igrad] : NULL;
 }
 
 struct lm_ggml_tensor * lm_ggml_graph_get_grad_acc(const struct lm_ggml_cgraph * cgraph, const struct lm_ggml_tensor * node) {
     const size_t igrad = lm_ggml_hash_find(&cgraph->visited_hash_set, node);
-    return igrad != LM_GGML_HASHSET_FULL && lm_ggml_bitset_get(cgraph->visited_hash_set.used, igrad) ? cgraph->grad_accs[igrad] : NULL;
+    return igrad != LM_GGML_HASHSET_FULL && lm_ggml_bitset_get(cgraph->visited_hash_set.used, igrad) && cgraph->grad_accs ? cgraph->grad_accs[igrad] : NULL;
 }
 
 void lm_ggml_graph_print(const struct lm_ggml_cgraph * cgraph) {
@@ -6502,7 +6536,7 @@ struct lm_gguf_context {
     void * data;
 };
 
-static size_t lm_gguf_type_size(enum lm_gguf_type type) {
+size_t lm_gguf_type_size(enum lm_gguf_type type) {
     LM_GGML_ASSERT(0 <= type && type < LM_GGUF_TYPE_COUNT);
     return LM_GGUF_TYPE_SIZE[type];
 }
@@ -6630,13 +6664,7 @@ struct lm_gguf_context * lm_gguf_init_empty(void) {
     return ctx;
 }
 
-struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gguf_init_params params) {
-    FILE * file = lm_ggml_fopen(fname, "rb");
-    if (!file) {
-        fprintf(stderr, "%s: failed to open '%s': '%s'\n", __func__, fname, strerror(errno));
-        return NULL;
-    }
-
+struct lm_gguf_context * lm_gguf_init_from_file_impl(FILE * file, struct lm_gguf_init_params params) {
     // offset from start of file
     size_t offset = 0;
 
@@ -6649,7 +6677,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
         for (uint32_t i = 0; i < sizeof(magic); i++) {
             if (magic[i] != LM_GGUF_MAGIC[i]) {
                 fprintf(stderr, "%s: invalid magic characters '%c%c%c%c'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
-                fclose(file);
                 return NULL;
             }
         }
@@ -6660,7 +6687,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
     struct lm_gguf_context * ctx = calloc(1, sizeof(struct lm_gguf_context));
     if (!ctx) {
         fprintf(stderr, "%s: failed to allocate memory for context\n", __func__);
-        fclose(file);
         return NULL;
     }
 
@@ -6678,7 +6704,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
 
         if (ctx->header.version == 1) {
             fprintf(stderr, "%s: GGUFv1 is no longer supported. please use a more up-to-date version\n", __func__);
-            fclose(file);
             lm_gguf_free(ctx);
             return NULL;
         }
@@ -6691,7 +6716,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
 
         if (!ok) {
             fprintf(stderr, "%s: failed to read header\n", __func__);
-            fclose(file);
             lm_gguf_free(ctx);
             return NULL;
         }
@@ -6701,12 +6725,13 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
     {
         const uint64_t n_kv = ctx->header.n_kv;
 
-        ctx->kv = calloc(n_kv, sizeof(struct lm_gguf_kv));
-        if (!ctx->kv) {
-            fprintf(stderr, "%s: failed to allocate memory for kv pairs\n", __func__);
-            fclose(file);
-            lm_gguf_free(ctx);
-            return NULL;
+        if (n_kv > 0) {
+            ctx->kv = calloc(n_kv, sizeof(struct lm_gguf_kv));
+            if (!ctx->kv) {
+                fprintf(stderr, "%s: failed to allocate memory for kv pairs\n", __func__);
+                lm_gguf_free(ctx);
+                return NULL;
+            }
         }
 
         for (uint64_t i = 0; i < n_kv; ++i) {
@@ -6753,7 +6778,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
                                     // prevent from integer overflow in the malloc below
                                     if (kv->value.arr.n >= SIZE_MAX/lm_gguf_type_size(kv->value.arr.type)) {
                                         fprintf(stderr, "%s: array size is too large (%" PRIu64 ")\n", __func__, kv->value.arr.n);
-                                        fclose(file);
                                         lm_gguf_free(ctx);
                                         return NULL;
                                     }
@@ -6761,7 +6785,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
                                     kv->value.arr.data = calloc(kv->value.arr.n, lm_gguf_type_size(kv->value.arr.type));
                                     if (!kv->value.arr.data) {
                                         fprintf(stderr, "%s: failed to allocate memory for array\n", __func__);
-                                        fclose(file);
                                         lm_gguf_free(ctx);
                                         return NULL;
                                     }
@@ -6773,7 +6796,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
                                     // prevent from integer overflow in the malloc below
                                     if (kv->value.arr.n >= SIZE_MAX/sizeof(struct lm_gguf_str)) {
                                         fprintf(stderr, "%s: array size is too large (%" PRIu64 ")\n", __func__, kv->value.arr.n);
-                                        fclose(file);
                                         lm_gguf_free(ctx);
                                         return NULL;
                                     }
@@ -6781,7 +6803,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
                                     kv->value.arr.data = calloc(kv->value.arr.n, sizeof(struct lm_gguf_str));
                                     if (!kv->value.arr.data) {
                                         fprintf(stderr, "%s: failed to allocate memory for array\n", __func__);
-                                        fclose(file);
                                         lm_gguf_free(ctx);
                                         return NULL;
                                     }
@@ -6812,7 +6833,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
 
         if (!ok) {
             fprintf(stderr, "%s: failed to read key-value pairs\n", __func__);
-            fclose(file);
             lm_gguf_free(ctx);
             return NULL;
         }
@@ -6823,7 +6843,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
         ctx->infos = calloc(ctx->header.n_tensors, sizeof(struct lm_gguf_tensor_info));
         if (!ctx->infos) {
             fprintf(stderr, "%s: failed to allocate memory for tensor infos\n", __func__);
-            fclose(file);
             lm_gguf_free(ctx);
             return NULL;
         }
@@ -6859,7 +6878,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
 
             if (!ok) {
                 fprintf(stderr, "%s: failed to read tensor info\n", __func__);
-                fclose(file);
                 lm_gguf_free(ctx);
                 return NULL;
             }
@@ -6902,7 +6920,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
                 // this tensor type support have been removed:
                 fprintf(stderr, "%s: tensor '%s' of type %d: %s\n",
                         __func__, info->name.data, (int) info->type, lm_ggml_type_name(info->type));
-                fclose(file);
                 lm_gguf_free(ctx);
                 return NULL;
             }
@@ -6910,7 +6927,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
             if (ne % lm_ggml_blck_size(info->type) != 0) {
                 fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
                         __func__, info->name.data, (int) info->type, lm_ggml_type_name(info->type), ne, lm_ggml_blck_size(info->type));
-                fclose(file);
                 lm_gguf_free(ctx);
                 return NULL;
             }
@@ -6942,7 +6958,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
         *params.ctx = lm_ggml_init(pdata);
         if (*params.ctx == NULL) {
             fprintf(stderr, "%s: failed to initialize context\n", __func__);
-            fclose(file);
             lm_gguf_free(ctx);
             return NULL;
         }
@@ -6961,7 +6976,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
 
             if (!ok) {
                 fprintf(stderr, "%s: failed to read tensor data\n", __func__);
-                fclose(file);
                 lm_ggml_free(ctx_data);
                 lm_gguf_free(ctx);
                 return NULL;
@@ -7000,7 +7014,6 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
 
         if (!ok) {
             fprintf(stderr, "%s: failed to read the tensor data\n", __func__);
-            fclose(file);
             lm_ggml_free(ctx_data);
             lm_gguf_free(ctx);
             return NULL;
@@ -7009,11 +7022,21 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
         lm_ggml_set_no_alloc(ctx_data, params.no_alloc);
     }
 
-    fclose(file);
-
     return ctx;
 }
 
+struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gguf_init_params params) {
+    FILE * file = lm_ggml_fopen(fname, "rb");
+    if (!file) {
+        fprintf(stderr, "%s: failed to open '%s': '%s'\n", __func__, fname, strerror(errno));
+        return NULL;
+    }
+
+    struct lm_gguf_context * result = lm_gguf_init_from_file_impl(file, params);
+    fclose(file);
+    return result;
+}
+
 void lm_gguf_free(struct lm_gguf_context * ctx) {
     if (ctx == NULL) {
         return;
@@ -7473,13 +7496,7 @@ void lm_gguf_set_tensor_data(struct lm_gguf_context * ctx, const char * name, co
 //    fwrite(val, sizeof(char), size, file);
 //}
 
-struct lm_gguf_buf {
-    void * data;
-    size_t size;
-    size_t offset;
-};
-
-static struct lm_gguf_buf lm_gguf_buf_init(size_t size) {
+struct lm_gguf_buf lm_gguf_buf_init(size_t size) {
     struct lm_gguf_buf buf = {
         /*buf.data   =*/ size == 0 ? NULL : LM_GGML_CALLOC(1, size),
         /*buf.size   =*/ size,
@@ -7489,7 +7506,7 @@ static struct lm_gguf_buf lm_gguf_buf_init(size_t size) {
     return buf;
 }
 
-static void lm_gguf_buf_free(struct lm_gguf_buf buf) {
+void lm_gguf_buf_free(struct lm_gguf_buf buf) {
     if (buf.data) {
         LM_GGML_FREE(buf.data);
     }
@@ -7527,7 +7544,7 @@ static void lm_gguf_bwrite_el(struct lm_gguf_buf * buf, const void * val, size_t
     buf->offset += el_size;
 }
 
-static void lm_gguf_write_to_buf(const struct lm_gguf_context * ctx, struct lm_gguf_buf * buf, bool only_meta) {
+void lm_gguf_write_to_buf(const struct lm_gguf_context * ctx, struct lm_gguf_buf * buf, bool only_meta) {
     // write header
     lm_gguf_bwrite_el(buf, &ctx->header.magic,     sizeof(ctx->header.magic));
     lm_gguf_bwrite_el(buf, &ctx->header.version,   sizeof(ctx->header.version));

package/cpp/ggml.h

@@ -1565,17 +1565,6 @@ extern "C" {
         int                   d1, // dilation dimension 1
         bool                  is_2D);
 
-    LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_depthwise_2d(
-            struct lm_ggml_context * ctx,
-            struct lm_ggml_tensor  * a,  // convolution kernel
-            struct lm_ggml_tensor  * b,  // data
-            int                  s0,  // stride dimension 0
-            int                  s1,  // stride dimension 1
-            int                  p0,  // padding dimension 0
-            int                  p1,  // padding dimension 1
-            int                  d0,  // dilation dimension 0
-            int                  d1); // dilation dimension 1
-
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_1d(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,   // convolution kernel
@@ -1593,6 +1582,23 @@ extern "C" {
             int                   s,  // stride
             int                   d); // dilation
 
+    // depthwise
+    // TODO: this is very likely wrong for some cases! - needs more testing
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_1d_dw(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a,   // convolution kernel
+            struct lm_ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
+
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_1d_dw_ph(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a,   // convolution kernel
+            struct lm_ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   d0); // dilation
+
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_transpose_1d(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,   // convolution kernel
@@ -1612,7 +1618,6 @@ extern "C" {
             int                   d0,  // dilation dimension 0
             int                   d1); // dilation dimension 1
 
-
     // kernel size is a->ne[0] x a->ne[1]
     // stride is equal to kernel size
     // padding is zero
@@ -1639,6 +1644,18 @@ extern "C" {
             struct lm_ggml_tensor  * a,
             struct lm_ggml_tensor  * b);
 
+    // depthwise
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_2d_dw(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a,  // convolution kernel
+            struct lm_ggml_tensor  * b,  // data
+            int                  s0,  // stride dimension 0
+            int                  s1,  // stride dimension 1
+            int                  p0,  // padding dimension 0
+            int                  p1,  // padding dimension 1
+            int                  d0,  // dilation dimension 0
+            int                  d1); // dilation dimension 1
+
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_conv_transpose_2d_p0(
             struct lm_ggml_context * ctx,
             struct lm_ggml_tensor  * a,

package/cpp/llama-grammar.cpp

@@ -822,15 +822,11 @@ llama_grammar_stacks & llama_grammar_get_stacks(struct llama_grammar * grammar)
     return grammar->stacks;
 }
 
-void llama_grammar_accept(
-        const llama_grammar_rules  & rules,
-        const llama_grammar_stacks & stacks,
-        const uint32_t               chr,
-              llama_grammar_stacks & stacks_new) {
-    stacks_new.clear();
-    stacks_new.reserve(stacks.size());
+void llama_grammar_accept(struct llama_grammar * grammar, uint32_t chr) {
+    llama_grammar_stacks stacks_new;
+    stacks_new.reserve(grammar->stacks.size());
 
-    for (const auto & stack : stacks) {
+    for (const auto & stack : grammar->stacks) {
         if (stack.empty()) {
             continue;
         }
@@ -844,9 +840,11 @@ void llama_grammar_accept(
             if (!llama_grammar_is_end_of_sequence(pos)) {
                 new_stack.push_back(pos);
             }
-            llama_grammar_advance_stack(rules, new_stack, stacks_new);
+            llama_grammar_advance_stack(grammar->rules, new_stack, stacks_new);
         }
     }
+
+    grammar->stacks = std::move(stacks_new);
 }
 
 llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
@@ -1051,7 +1049,12 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) {
 }
 
 struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar) {
-    llama_grammar * result = new llama_grammar { grammar.vocab, grammar.rules, grammar.stacks, grammar.partial_utf8, };
+    llama_grammar * result = new llama_grammar {
+        grammar.vocab,
+        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++) {
@@ -1059,7 +1062,7 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
             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];
+                        result->stacks[is][ie] =  &result->rules[ir0][ir1];
                     }
                 }
             }
@@ -1126,11 +1129,8 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
     const auto   decoded     = decode_utf8(piece, grammar.partial_utf8);
     const auto & code_points = decoded.first;
 
-    llama_grammar_stacks stacks_new;
-
     for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
-        llama_grammar_accept(grammar.rules, grammar.stacks, *it, stacks_new);
-        grammar.stacks = std::move(stacks_new);
+        llama_grammar_accept(&grammar, *it);
     }
 
     grammar.partial_utf8 = decoded.second;