whispercpp 1.3.6 → 1.3.7

data/ext/sources/ggml/src/ggml.c

@@ -53,6 +53,21 @@
 
 #define UNUSED GGML_UNUSED
 
+uint64_t ggml_graph_next_uid(void) {
+#ifdef _MSC_VER
+#if defined(_WIN32)
+    static volatile LONG counter = 1;
+    return (uint64_t) InterlockedIncrement(&counter) - 1;
+#else
+    static volatile long long counter = 1;
+    return (uint64_t) _InterlockedIncrement64(&counter) - 1;
+#endif
+#else
+    static uint64_t counter = 1;
+    return __atomic_fetch_add(&counter, 1, __ATOMIC_RELAXED);
+#endif
+}
+
 // Needed for ggml_fp32_to_bf16_row()
 #if defined(__AVX512BF16__)
 #if defined(_MSC_VER)
@@ -651,6 +666,14 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .to_float                 = (ggml_to_float_t) ggml_fp16_to_fp32_row,
         .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_fp16_row,
     },
+    [GGML_TYPE_Q1_0] = {
+        .type_name                = "q1_0",
+        .blck_size                = QK1_0,
+        .type_size                = sizeof(block_q1_0),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_q1_0,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q1_0_ref,
+    },
     [GGML_TYPE_Q4_0] = {
         .type_name                = "q4_0",
         .blck_size                = QK4_0,
@@ -1008,6 +1031,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "IM2COL",
     "IM2COL_BACK",
     "IM2COL_3D",
+    "COL2IM_1D",
     "CONV_2D",
     "CONV_3D",
     "CONV_2D_DW",
@@ -1057,7 +1081,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 96, "GGML_OP_COUNT != 96");
+static_assert(GGML_OP_COUNT == 97, "GGML_OP_COUNT != 97");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1118,6 +1142,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "im2col(x)",
     "im2col_back(x)",
     "im2col_3d(x)",
+    "col2im_1d(x)",
     "conv_2d(x)",
     "conv_3d(x)",
     "conv_2d_dw(x)",
@@ -1167,7 +1192,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 96, "GGML_OP_COUNT != 96");
+static_assert(GGML_OP_COUNT == 97, "GGML_OP_COUNT != 97");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -1384,6 +1409,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_BF16:          wtype = GGML_TYPE_BF16;  break;
         case GGML_FTYPE_MOSTLY_Q4_0:          wtype = GGML_TYPE_Q4_0;  break;
         case GGML_FTYPE_MOSTLY_Q4_1:          wtype = GGML_TYPE_Q4_1;  break;
+        case GGML_FTYPE_MOSTLY_Q1_0:          wtype = GGML_TYPE_Q1_0;  break;
         case GGML_FTYPE_MOSTLY_Q5_0:          wtype = GGML_TYPE_Q5_0;  break;
         case GGML_FTYPE_MOSTLY_Q5_1:          wtype = GGML_TYPE_Q5_1;  break;
         case GGML_FTYPE_MOSTLY_Q8_0:          wtype = GGML_TYPE_Q8_0;  break;
@@ -3240,6 +3266,16 @@ void ggml_mul_mat_set_prec(
     ggml_set_op_params_i32(a, 0, prec_i32);
 }
 
+void ggml_mul_mat_set_hint(
+        struct ggml_tensor * a,
+        enum ggml_op_hint    hint) {
+    GGML_ASSERT(a->op == GGML_OP_MUL_MAT);
+
+    const int32_t hint_i32 = (int32_t) hint;
+
+    ggml_set_op_params_i32(a, 1, hint_i32);
+}
+
 // ggml_mul_mat_id
 
 /*
@@ -4507,6 +4543,41 @@ struct ggml_tensor * ggml_conv_1d_dw_ph(
     return ggml_conv_1d_dw(ctx, a, b, s0, a->ne[0] / 2, d0);
 }
 
+// ggml_col2im_1d
+
+struct ggml_tensor * ggml_col2im_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   s0,
+        int                   oc,
+        int                   p0) {
+    GGML_ASSERT(ggml_is_matrix(a));
+    GGML_ASSERT(ggml_is_contiguous(a));
+    GGML_ASSERT(a->type == GGML_TYPE_F32 || a->type == GGML_TYPE_F16 || a->type == GGML_TYPE_BF16);
+    GGML_ASSERT(s0 > 0);
+    GGML_ASSERT(oc > 0);
+    GGML_ASSERT(p0 >= 0);
+
+    const int64_t K_OC = a->ne[0];
+    const int64_t T_in = a->ne[1];
+    const int64_t K = K_OC / oc;
+    const int64_t T_out = (T_in - 1) * s0 + K - 2 * p0;
+
+    GGML_ASSERT(K_OC == K * oc);  // a->ne[0] must be a whole number of oc blocks
+    GGML_ASSERT(K > 0 && T_out > 0);
+
+    const int64_t ne[4] = { T_out, oc, 1, 1 };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, 2, ne);
+
+    int32_t params[] = { s0, (int32_t)oc, (int32_t)p0 };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_COL2IM_1D;
+    result->src[0] = a;
+
+    return result;
+}
+
 // ggml_conv_transpose_1d
 
 static int64_t ggml_calc_conv_transpose_1d_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
@@ -4962,6 +5033,7 @@ static struct ggml_tensor * ggml_interpolate_impl(
     GGML_ASSERT((mode & 0xFF) < GGML_SCALE_MODE_COUNT);
     // TODO: implement antialias for modes other than bilinear
     GGML_ASSERT(!(mode & GGML_SCALE_FLAG_ANTIALIAS) || (mode & 0xFF) == GGML_SCALE_MODE_BILINEAR);
+    GGML_ASSERT(a->type == GGML_TYPE_F32);
 
     struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
 
@@ -5188,7 +5260,7 @@ static struct ggml_tensor * ggml_fill_impl(
     struct ggml_tensor  * a,
     float                 c,
     bool                  inplace) {
-    GGML_ASSERT(a->type == GGML_TYPE_F32);
+    GGML_ASSERT(a->type == GGML_TYPE_F32 || a->type == GGML_TYPE_F16);
     GGML_ASSERT(ggml_is_contiguous(a));
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
@@ -5307,6 +5379,7 @@ struct ggml_tensor * ggml_flash_attn_ext(
     GGML_ASSERT(q->ne[3] == v->ne[3]);
 
     if (mask) {
+        GGML_ASSERT(mask->type == GGML_TYPE_F16);
         GGML_ASSERT(ggml_is_contiguous(mask));
         //GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
 
@@ -6150,7 +6223,8 @@ struct ggml_tensor * ggml_gated_delta_net(
         struct ggml_tensor  * v,
         struct ggml_tensor  * g,
         struct ggml_tensor  * beta,
-        struct ggml_tensor  * state) {
+        struct ggml_tensor  * state,
+        int64_t               K) {
     GGML_ASSERT(ggml_is_contiguous_rows(q));
     GGML_ASSERT(ggml_is_contiguous_rows(k));
     GGML_ASSERT(ggml_is_contiguous_rows(v));
@@ -6174,13 +6248,18 @@ struct ggml_tensor * ggml_gated_delta_net(
     GGML_ASSERT(g->ne[0] == 1 || g->ne[0] == S_v);
     GGML_ASSERT(beta->ne[0] == 1);
 
-    GGML_ASSERT(ggml_nelements(state) == S_v * S_v * H * n_seqs);
-
-    // concat output and new_state into a single tensor
-    // output: S_v * H * n_tokens * n_seqs, state: S_v * S_v * H * n_seqs
-    const int64_t ne[4] = { S_v * H, n_tokens * n_seqs + S_v * n_seqs, 1, 1 };
+    // state holds the initial state s0 only: [S_v, S_v, H, n_seqs]. K (snapshot slot count) is an op param.
+    GGML_ASSERT(state->ne[0] == S_v);
+    GGML_ASSERT(state->ne[1] == S_v);
+    GGML_ASSERT(state->ne[2] == H);
+    GGML_ASSERT(state->ne[3] == n_seqs);
+    GGML_ASSERT(K >= 1);
+    const int64_t state_rows = K * S_v * n_seqs;
+    const int64_t ne[4] = { S_v * H, n_tokens * n_seqs + state_rows, 1, 1 };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
+    ggml_set_op_params_i32(result, 0, (int32_t) K);
+
     result->op     = GGML_OP_GATED_DELTA_NET;
     result->src[0] = q;
     result->src[1] = k;
@@ -7087,6 +7166,7 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
         /*.use_counts   =*/ use_counts_ptr,
         /*.hash_table   =*/ { hash_size, hash_used, hash_keys_ptr },
         /*.order        =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
+        /*.uid          =*/ 0,
     };
 
     ggml_hash_set_reset(&cgraph->visited_hash_set);
@@ -7114,6 +7194,7 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1)
         /*.use_counts       =*/ cgraph0->use_counts,
         /*.visited_hash_set =*/ cgraph0->visited_hash_set,
         /*.order            =*/ cgraph0->order,
+        /*.uid              =*/ 0
     };
 
     return cgraph;
@@ -7633,7 +7714,7 @@ size_t ggml_quantize_chunk(
                int64_t   nrows,
                int64_t   n_per_row,
            const float * imatrix) {
-    const int64_t n = (int64_t) nrows * n_per_row;
+    const int64_t n = nrows * n_per_row;
 
     if (ggml_quantize_requires_imatrix(type)) {
         GGML_ASSERT(imatrix != NULL);
@@ -7650,20 +7731,21 @@ size_t ggml_quantize_chunk(
     size_t result = 0;
 
     switch (type) {
-        case GGML_TYPE_Q4_0:    result = quantize_q4_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q4_1:    result = quantize_q4_1(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q5_0:    result = quantize_q5_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q5_1:    result = quantize_q5_1(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q8_0:    result = quantize_q8_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_MXFP4:   result = quantize_mxfp4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_NVFP4:   result = quantize_nvfp4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q2_K:    result = quantize_q2_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q3_K:    result = quantize_q3_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q4_K:    result = quantize_q4_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q5_K:    result = quantize_q5_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q6_K:    result = quantize_q6_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_TQ1_0:   result = quantize_tq1_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_TQ2_0:   result = quantize_tq2_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q1_0:    result = quantize_q1_0   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_0:    result = quantize_q4_0   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_1:    result = quantize_q4_1   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q5_0:    result = quantize_q5_0   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q5_1:    result = quantize_q5_1   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q8_0:    result = quantize_q8_0   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_MXFP4:   result = quantize_mxfp4  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_NVFP4:   result = quantize_nvfp4  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q2_K:    result = quantize_q2_K   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q3_K:    result = quantize_q3_K   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_K:    result = quantize_q4_K   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q5_K:    result = quantize_q5_K   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q6_K:    result = quantize_q6_K   (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_TQ1_0:   result = quantize_tq1_0  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_TQ2_0:   result = quantize_tq2_0  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_XXS: result = quantize_iq2_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_XS:  result = quantize_iq2_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ3_XXS: result = quantize_iq3_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
@@ -7728,9 +7810,9 @@ struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) {
 }
 
 bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) {
-    if (p0->n_threads      != p1->n_threads  )    return false;
-    if (p0->prio           != p1->prio       )    return false;
-    if (p0->poll           != p1->poll       )    return false;
-    if (p0->strict_cpu     != p1->strict_cpu )    return false;
+    if (p0->n_threads  != p1->n_threads  ) return false;
+    if (p0->prio       != p1->prio       ) return false;
+    if (p0->poll       != p1->poll       ) return false;
+    if (p0->strict_cpu != p1->strict_cpu ) return false;
     return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0;
 }

data/ext/sources/ggml/src/gguf.cpp

@@ -228,9 +228,18 @@ struct gguf_context {
 };
 
 struct gguf_reader {
-    gguf_reader(FILE * file) : file(file) {
-        // read the remaining bytes once and update on each read
-        nbytes_remain = file_remain(file);
+    gguf_reader(
+            gguf_reader_callback_t callback,
+            void * userdata,
+            size_t max_chunk_read,
+            uint64_t data_offset = 0,
+            uint64_t nbytes_remain = 0)
+        : callback(callback),
+          userdata(userdata),
+          max_chunk_read(max_chunk_read),
+          data_offset(data_offset),
+          nbytes_remain(nbytes_remain) {
+        GGML_ASSERT(max_chunk_read > 0);
     }
 
     // helper for remaining bytes in a file
@@ -257,12 +266,10 @@ struct gguf_reader {
     template <typename T>
     bool read(T & dst) const {
         const size_t size = sizeof(dst);
-        if (nbytes_remain < size) {
+        if (size > nbytes_remain) {
             return false;
         }
-        const size_t nread = fread(&dst, 1, size, file);
-        nbytes_remain -= nread;
-        return nread == size;
+        return read_raw(&dst, size) == size;
     }
 
     template <typename T>
@@ -344,24 +351,71 @@ struct gguf_reader {
             return false;
         }
         dst.resize(static_cast<size_t>(size));
-        const size_t nread = fread(dst.data(), 1, size, file);
-        nbytes_remain -= nread;
-        return nread == size;
+        return read_raw(dst.data(), static_cast<size_t>(size)) == size;
     }
 
     bool read(void * dst, const size_t size) const {
         if (size > nbytes_remain) {
             return false;
         }
-        const size_t nread = fread(dst, 1, size, file);
-        nbytes_remain -= nread;
-        return nread == size;
+        return read_raw(dst, size) == size;
+    }
+
+    uint64_t tell() const {
+        return data_offset;
+    }
+
+    bool seek(uint64_t absolute_offset) const {
+        const uint64_t end_offset = uint64_t(data_offset) + nbytes_remain;
+        if (absolute_offset > end_offset) {
+            return false;
+        }
+
+        data_offset = absolute_offset;
+        nbytes_remain = end_offset - absolute_offset;
+
+        return true;
     }
 
 private:
-    FILE * file;
+    size_t read_raw(void * dst, size_t size) const {
+        if (callback == nullptr || size == 0) {
+            return 0;
+        }
+
+        uint8_t * data = static_cast<uint8_t *>(dst);
+        size_t total_nread = 0;
+        bool reached_eof = false;
+
+        while (total_nread < size) {
+            const size_t chunk_size = std::min(max_chunk_read, size - total_nread);
+            if (data_offset + total_nread < data_offset) {
+                break;
+            }
+            const size_t nread = callback(userdata, static_cast<void *>(data + total_nread), data_offset + total_nread, chunk_size);
+            total_nread += nread;
+            if (nread != chunk_size) {
+                reached_eof = true;
+                break;
+            }
+        }
 
-    mutable uint64_t nbytes_remain;
+        data_offset += total_nread;
+        GGML_ASSERT(total_nread <= nbytes_remain);
+        nbytes_remain -= total_nread;
+
+        if (reached_eof) {
+            nbytes_remain = 0;
+        }
+
+        return total_nread;
+    }
+
+    gguf_reader_callback_t callback = nullptr;
+    void * userdata = nullptr;
+    size_t max_chunk_read = 0;
+    mutable uint64_t data_offset = 0;
+    mutable uint64_t nbytes_remain = 0;
 };
 
 struct gguf_context * gguf_init_empty(void) {
@@ -394,8 +448,7 @@ bool gguf_read_emplace_helper(const struct gguf_reader & gr, std::vector<struct
     return true;
 }
 
-struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params) {
-    const struct gguf_reader gr(file);
+static struct gguf_context * gguf_init_from_reader(const struct gguf_reader & gr, struct gguf_init_params params) {
     struct gguf_context * ctx = new gguf_context;
 
     bool ok = true;
@@ -696,14 +749,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     GGML_ASSERT(int64_t(ctx->info.size()) == n_tensors);
 
     // we require the data section to be aligned, so take into account any padding
-    if (gguf_fseek(file, GGML_PAD(gguf_ftell(file), ctx->alignment), SEEK_SET) != 0) {
+    if (n_tensors > 0 && !gr.seek(GGML_PAD(gr.tell(), ctx->alignment))) {
         GGML_LOG_ERROR("%s: failed to seek to beginning of data section\n", __func__);
         gguf_free(ctx);
         return nullptr;
     }
 
     // store the current file offset - this is where the data section starts
-    ctx->offset = gguf_ftell(file);
+    ctx->offset = gr.tell();
 
     // compute the total size of the data section, taking into account the alignment
     {
@@ -840,6 +893,89 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     return ctx;
 }
 
+struct gguf_context * gguf_init_from_callback(gguf_reader_callback_t callback, void * userdata, size_t max_chunk_read, uint64_t max_expected_size, struct gguf_init_params params) {
+    if (callback == nullptr) {
+        return nullptr;
+    }
+
+    const struct gguf_reader gr(callback, userdata, max_chunk_read == 0 ? SIZE_MAX : max_chunk_read, 0, max_expected_size);
+    return gguf_init_from_reader(gr, params);
+}
+
+struct gguf_file_reader {
+    FILE * file;
+    uint64_t offset;
+};
+
+static size_t gguf_file_reader_callback(void * userdata, void * output, uint64_t offset, size_t len) {
+    GGML_ASSERT(len > 0);
+
+    gguf_file_reader & reader = *static_cast<gguf_file_reader *>(userdata);
+
+    if (reader.offset != offset) {
+        if (offset > INT64_MAX || gguf_fseek(reader.file, static_cast<int64_t>(offset), SEEK_SET) != 0) {
+            return 0;
+        }
+
+        reader.offset = offset;
+    }
+
+    const size_t nread = fread(static_cast<uint8_t *>(output), 1, len, reader.file);
+    reader.offset += nread;
+    return nread;
+}
+
+struct gguf_context * gguf_init_from_file_ptr(FILE * file, struct gguf_init_params params) {
+    if (!file) {
+        return nullptr;
+    }
+
+    const int64_t cur = gguf_ftell(file);
+    if (cur < 0) {
+        return nullptr;
+    }
+
+    gguf_file_reader reader = {
+        /*.file   = */ file,
+        /*.offset = */ static_cast<uint64_t>(cur),
+    };
+    const struct gguf_reader gr(gguf_file_reader_callback, &reader, SIZE_MAX, reader.offset, gguf_reader::file_remain(file));
+    return gguf_init_from_reader(gr, params);
+}
+
+struct gguf_buffer_reader {
+    const uint8_t * data;
+    size_t          size;
+};
+
+static size_t gguf_buffer_reader_callback(void * userdata, void * output, uint64_t offset, size_t len) {
+    GGML_ASSERT(len > 0);
+
+    const gguf_buffer_reader & reader = *static_cast<gguf_buffer_reader *>(userdata);
+
+    if (offset > reader.size || len > reader.size - offset) {
+        return 0;
+    }
+
+    const size_t data_offset = static_cast<size_t>(offset);
+    const size_t nread = std::min(len, reader.size - data_offset);
+    memcpy(static_cast<uint8_t *>(output), reader.data + data_offset, nread);
+    return nread;
+}
+
+struct gguf_context * gguf_init_from_buffer(const void * data, size_t size, struct gguf_init_params params) {
+    if (data == nullptr || size == 0) {
+        return nullptr;
+    }
+
+    gguf_buffer_reader reader = {
+        /*.data = */ static_cast<const uint8_t *>(data),
+        /*.size = */ size,
+    };
+    const struct gguf_reader gr(gguf_buffer_reader_callback, &reader, SIZE_MAX, 0, size);
+    return gguf_init_from_reader(gr, params);
+}
+
 struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params) {
     FILE * file = ggml_fopen(fname, "rb");
 
@@ -848,7 +984,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
         return nullptr;
     }
 
-    struct gguf_context * result = gguf_init_from_file_impl(file, params);
+    struct gguf_context * result = gguf_init_from_file_ptr(file, params);
     fclose(file);
     return result;
 }
@@ -1508,6 +1644,19 @@ void gguf_write_to_buf(const struct gguf_context * ctx, std::vector<int8_t> & bu
     gguf_write_out(ctx, gw, only_meta);
 }
 
+bool gguf_write_to_file_ptr(const struct gguf_context * ctx, FILE * file, bool only_meta) {
+    GGML_ASSERT(file);
+
+    try {
+        gguf_writer_file gw(file);
+        gguf_write_out(ctx, gw, only_meta);
+    } catch (const std::runtime_error& ex) {
+        GGML_LOG_ERROR("%s: failed to write GGUF data: %s\n", __func__, ex.what());
+        return false;
+    }
+    return true;
+}
+
 bool gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta) {
     FILE * file = ggml_fopen(fname, "wb");
 
@@ -1516,17 +1665,13 @@ bool gguf_write_to_file(const struct gguf_context * ctx, const char * fname, boo
         return false;
     }
 
-    try {
-        gguf_writer_file gw(file);
-        gguf_write_out(ctx, gw, only_meta);
-    } catch (const std::runtime_error& ex) {
-        GGML_LOG_ERROR("%s: failed to write GGUF data into '%s': %s\n", __func__, fname, ex.what());
-        fclose(file);
-        return false;
+    const bool success = gguf_write_to_file_ptr(ctx, file, only_meta);
+    if (!success) {
+        GGML_LOG_ERROR("%s: failed to write GGUF data into '%s'\n", __func__, fname);
     }
 
     fclose(file);
-    return true;
+    return success;
 }
 
 size_t gguf_get_meta_size(const struct gguf_context * ctx) {