gguf : fix ftell/fseek for Windows (#19870)

* server: fix query params lost when proxying requests in multi-model router mode

* server: re-encode query params using httplib::encode_query_component in proxy

ggml/include/ggml.h

@@ -730,10 +730,6 @@ extern "C" {
     GGML_API size_t  ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
     GGML_API size_t  ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
 
-    GGML_DEPRECATED(
-    GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
-    "use ggml_row_size() instead");
-
     GGML_API const char * ggml_type_name(enum ggml_type type);
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);

ggml/src/ggml.c

@@ -899,7 +899,8 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
 };
 
 const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type) {
-    GGML_ASSERT(type < GGML_TYPE_COUNT);
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
     return &type_traits[type];
 }
 
@@ -1265,27 +1266,33 @@ size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
 }
 
 int64_t ggml_blck_size(enum ggml_type type) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
     return type_traits[type].blck_size;
 }
 
 size_t ggml_type_size(enum ggml_type type) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
     return type_traits[type].type_size;
 }
 
 size_t ggml_row_size(enum ggml_type type, int64_t ne) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
     assert(ne % ggml_blck_size(type) == 0);
     return ggml_type_size(type)*ne/ggml_blck_size(type);
 }
 
-double ggml_type_sizef(enum ggml_type type) {
-    return ((double)(type_traits[type].type_size))/type_traits[type].blck_size;
-}
-
 const char * ggml_type_name(enum ggml_type type) {
-    return type < GGML_TYPE_COUNT ? type_traits[type].type_name : "NONE";
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
+    return type_traits[type].type_name;
 }
 
 bool ggml_is_quantized(enum ggml_type type) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
     return type_traits[type].is_quantized;
 }
 
@@ -1629,11 +1636,23 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
     const size_t cur_end  = cur_offs + cur_size;
 
     // align to GGML_MEM_ALIGN
+    GGML_ASSERT(size <= SIZE_MAX - (GGML_MEM_ALIGN - 1));
     size_t size_needed = GGML_PAD(size, GGML_MEM_ALIGN);
 
     char * const mem_buffer = ctx->mem_buffer;
     struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end);
 
+    // integer overflow checks
+    if (cur_end > SIZE_MAX - size_needed) {
+        GGML_LOG_WARN("%s: overflow detected in cur_end (%zu) + size_needed (%zu)\n", __func__, cur_end, size_needed);
+        return NULL;
+    }
+    if (cur_end + size_needed > SIZE_MAX - GGML_OBJECT_SIZE) {
+        GGML_LOG_WARN("%s: overflow detected in cur_end (%zu) + size_needed (%zu) + GGML_OBJECT_SIZE (%zu)\n", __func__,
+                cur_end, size_needed, (size_t) GGML_OBJECT_SIZE);
+        return NULL;
+    }
+
     if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
         GGML_LOG_WARN("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
                 __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size);
@@ -1702,6 +1721,8 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         obj_alloc_size = data_size;
     }
 
+    GGML_ASSERT(GGML_TENSOR_SIZE <= SIZE_MAX - obj_alloc_size);
+
     struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TYPE_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size);
     GGML_ASSERT(obj_new);
 

ggml/src/gguf.cpp

@@ -15,6 +15,17 @@
 #include <string>
 #include <vector>
 
+#define GGUF_MAX_STRING_LENGTH  (1024*1024*1024)
+#define GGUF_MAX_ARRAY_ELEMENTS (1024*1024*1024)
+
+#ifdef _WIN32
+#    define gguf_ftell _ftelli64
+#    define gguf_fseek _fseeki64
+#else
+#    define gguf_ftell ftello
+#    define gguf_fseek fseeko
+#endif
+
 template <typename T>
 struct type_to_gguf_type;
 
@@ -228,6 +239,26 @@ struct gguf_reader {
 
     template <typename T>
     bool read(std::vector<T> & dst, const size_t n) const {
+        if (n > GGUF_MAX_ARRAY_ELEMENTS) {
+            return false;
+        }
+        const uint64_t nbytes = nbytes_remain();
+        if constexpr (std::is_same<T, std::string>::value) {
+            // strings are prefixed with their length, so we need to account for that
+            if (n > SIZE_MAX / sizeof(uint64_t)) {
+                return false;
+            }
+            if (nbytes < n * sizeof(uint64_t)) {
+                return false;
+            }
+        } else {
+            if (n > SIZE_MAX / sizeof(T)) {
+                return false;
+            }
+            if (nbytes < n * sizeof(T)) {
+                return false;
+            }
+        }
         dst.resize(n);
         for (size_t i = 0; i < dst.size(); ++i) {
             if constexpr (std::is_same<T, bool>::value) {
@@ -277,13 +308,43 @@ struct gguf_reader {
         if (!read(size)) {
             return false;
         }
-        dst.resize(size);
+        if (size > GGUF_MAX_STRING_LENGTH) {
+            GGML_LOG_ERROR("%s: string length %" PRIu64 " exceeds maximum %" PRIu64 "\n", __func__, size, (uint64_t) GGUF_MAX_STRING_LENGTH);
+            return false;
+        }
+        const uint64_t nbytes = nbytes_remain();
+        if (size > nbytes) {
+            GGML_LOG_ERROR("%s: string length %" PRIu64 " exceeds remaining file size %" PRIu64 " bytes\n", __func__, size, nbytes);
+            return false;
+        }
+        dst.resize(static_cast<size_t>(size));
         return fread(dst.data(), 1, dst.length(), file) == dst.length();
     }
 
     bool read(void * dst, const size_t size) const {
         return fread(dst, 1, size, file) == size;
     }
+
+    // remaining bytes in the file
+    uint64_t nbytes_remain() const {
+        const int64_t cur = gguf_ftell(file);
+        if (cur < 0) {
+            return 0;
+        }
+        if (gguf_fseek(file, 0, SEEK_END) != 0) {
+            gguf_fseek(file, cur, SEEK_SET);
+
+            return 0;
+        }
+        const int64_t end = gguf_ftell(file);
+        if (end < 0) {
+            gguf_fseek(file, cur, SEEK_SET);
+
+            return 0;
+        }
+        gguf_fseek(file, cur, SEEK_SET);
+        return static_cast<uint64_t>(end - cur);
+    }
 };
 
 struct gguf_context * gguf_init_empty(void) {
@@ -568,8 +629,8 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
             // check that tensor type is within defined range
             if (info.t.type < 0 || info.t.type >= GGML_TYPE_COUNT) {
-                GGML_LOG_ERROR("%s: tensor '%s' has invalid ggml type %d (%s)\n",
-                    __func__, info.t.name, info.t.type, ggml_type_name(info.t.type));
+                GGML_LOG_ERROR("%s: tensor '%s' has invalid ggml type %d. should be in [0, %d)\n",
+                    __func__, info.t.name, info.t.type, GGML_TYPE_COUNT);
                 ok = false;
                 break;
             }
@@ -618,14 +679,14 @@ 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 (fseek(file, GGML_PAD(ftell(file), ctx->alignment), SEEK_SET) != 0) {
+    if (gguf_fseek(file, GGML_PAD(gguf_ftell(file), ctx->alignment), SEEK_SET) != 0) {
         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 = ftell(file);
+    ctx->offset = gguf_ftell(file);
 
     // compute the total size of the data section, taking into account the alignment
     {
@@ -657,10 +718,34 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
         //   the ggml_tensor structs to the appropriate locations in the binary blob
 
         // compute the exact size needed for the new ggml_context
-        const size_t mem_size =
-            params.no_alloc ?
-            (n_tensors    )*ggml_tensor_overhead() :
-            (n_tensors + 1)*ggml_tensor_overhead() + ctx->size;
+        size_t mem_size = 0;
+        if (params.no_alloc) {
+            if (n_tensors != 0 && SIZE_MAX / n_tensors < ggml_tensor_overhead()) {
+                GGML_LOG_ERROR("%s: memory size overflow while allocating ggml context\n", __func__);
+                gguf_free(ctx);
+                return nullptr;
+            }
+
+            const size_t overhead = n_tensors * ggml_tensor_overhead();
+
+            mem_size = overhead;
+        } else {
+            if ((n_tensors + 1) != 0 && SIZE_MAX / (n_tensors + 1) < ggml_tensor_overhead()) {
+                GGML_LOG_ERROR("%s: memory size overflow while allocating ggml context\n", __func__);
+                gguf_free(ctx);
+                return nullptr;
+            }
+
+            const size_t overhead = (n_tensors + 1) * ggml_tensor_overhead();
+
+            if (SIZE_MAX - overhead < ctx->size) {
+                GGML_LOG_ERROR("%s: memory size overflow while allocating ggml context\n", __func__);
+                gguf_free(ctx);
+                return nullptr;
+            }
+
+            mem_size = overhead + ctx->size;
+        }
 
         struct ggml_init_params pdata = {
             /*mem_size   =*/ mem_size,

gguf-py/gguf/gguf_reader.py

@@ -175,6 +175,9 @@ class GGUFReader:
             if new_align.types != [GGUFValueType.UINT32]:
                 raise ValueError('Bad type for general.alignment field')
             self.alignment = new_align.parts[-1][0]
+            # Ensure alignment is a non-zero power of two
+            if self.alignment == 0 or (self.alignment & (self.alignment - 1)) != 0:
+                raise ValueError('Invalid alignment: must be a non-zero power of two')
         padding = offs % self.alignment
         if padding != 0:
             offs += self.alignment - padding
@@ -202,11 +205,11 @@ class GGUFReader:
 
     def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int:
         if field.name in self.fields:
-            # TODO: add option to generate error on duplicate keys
-            # raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')
+            # TODO: add option to make this a warning and accept duplicate keys like below
+            raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')
 
-            logger.warning(f'Duplicate key {field.name} at offset {field.offset}')
-            self.fields[field.name + '_{}'.format(field.offset)] = field
+            # logger.warning(f'Duplicate key {field.name} at offset {field.offset}')
+            # self.fields[field.name + '_{}'.format(field.offset)] = field
         else:
             self.fields[field.name] = field
         return 0 if skip_sum else sum(int(part.nbytes) for part in field.parts)

gguf-py/gguf/gguf_writer.py

@@ -501,6 +501,8 @@ class GGUFWriter:
         self.add_uint32(Keys.General.QUANTIZATION_VERSION, quantization_version)
 
     def add_custom_alignment(self, alignment: int) -> None:
+        if alignment <= 0 or (alignment & (alignment - 1)) != 0:
+            raise ValueError('Invalid alignment: must be a non-zero power of two')
         self.data_alignment = alignment
         self.add_uint32(Keys.General.ALIGNMENT, alignment)
 

src/llama-model.cpp

@@ -123,6 +123,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_8B_A1B:        return "8B.A1B";
         case LLM_TYPE_16B_A1B:       return "16B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
+        case LLM_TYPE_24B_A2B:       return "24B.A2B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
         case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
         case LLM_TYPE_35B_A3B:       return "35B.A3B";
@@ -2381,7 +2382,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
                 }
 
-                type = LLM_TYPE_8B_A1B;
+                switch (hparams.n_layer) {
+                    case 24: type = LLM_TYPE_8B_A1B;  break;
+                    case 40: type = LLM_TYPE_24B_A2B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
             } break;
         case LLM_ARCH_SMALLTHINKER:
             {

src/llama-model.h

@@ -116,6 +116,7 @@ enum llm_type {
     LLM_TYPE_8B_A1B, // lfm2moe
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
+    LLM_TYPE_24B_A2B, // lfm2moe
     LLM_TYPE_30B_A3B,
     LLM_TYPE_31B_A3_5B,
     LLM_TYPE_35B_A3B, // Qwen3.5

src/models/kimi-linear.cpp

@@ -116,6 +116,8 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
         cur = build_norm(inpL, layer.attn_norm, NULL, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Check layer type by checking which tensors exist
         // KDA layers have ssm_a_log tensor, MLA layers have wkv_a_mqa tensor
         bool is_kda = (layer.ssm_a != nullptr);

src/models/qwen35.cpp

@@ -29,6 +29,8 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
@@ -269,7 +271,6 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     cb(state_update_target, "state_update_target", il);
 
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
-    cb(conv_states_all, "conv_states_updated", il);
 
     ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
     state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);

src/models/qwen35moe.cpp

@@ -29,6 +29,8 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
@@ -269,7 +271,6 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     cb(state_update_target, "state_update_target", il);
 
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
-    cb(conv_states_all, "conv_states_updated", il);
 
     ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
     state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);

src/models/qwen3next.cpp

@@ -21,6 +21,8 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
@@ -354,7 +356,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     cb(state_update_target, "state_update_target", il);
 
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
-    cb(conv_states_all, "conv_states_updated", il);
 
     ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
     state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);

tests/test-gguf.cpp

@@ -48,6 +48,7 @@ enum handcrafted_file_type {
     HANDCRAFTED_DATA_NOT_ENOUGH_DATA       =  10 + offset_has_data,
     HANDCRAFTED_DATA_BAD_ALIGN             =  15 + offset_has_data,
     HANDCRAFTED_DATA_INCONSISTENT_ALIGN    =  20 + offset_has_data,
+    HANDCRAFTED_DATA_MEM_SIZE_OVERFLOW     =  30 + offset_has_data,
     HANDCRAFTED_DATA_SUCCESS               = 800 + offset_has_data,
     HANDCRAFTED_DATA_CUSTOM_ALIGN          = 810 + offset_has_data,
 };
@@ -84,6 +85,7 @@ static std::string handcrafted_file_type_name(const enum handcrafted_file_type h
         case HANDCRAFTED_DATA_NOT_ENOUGH_DATA:       return "DATA_NOT_ENOUGH_DATA";
         case HANDCRAFTED_DATA_BAD_ALIGN:             return "DATA_BAD_ALIGN";
         case HANDCRAFTED_DATA_INCONSISTENT_ALIGN:    return "DATA_INCONSISTENT_ALIGN";
+        case HANDCRAFTED_DATA_MEM_SIZE_OVERFLOW:     return "DATA_MEM_SIZE_OVERFLOW";
         case HANDCRAFTED_DATA_SUCCESS:               return "DATA_SUCCESS";
         case HANDCRAFTED_DATA_CUSTOM_ALIGN:          return "DATA_CUSTOM_ALIGN";
     }
@@ -196,6 +198,13 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         tensor_configs = get_tensor_configs(rng);
     }
 
+    if (hft == HANDCRAFTED_DATA_MEM_SIZE_OVERFLOW) {
+        tensor_configs.resize(2);
+
+        tensor_configs[0] = { GGML_TYPE_I8, { 0x7FFFFFFFFFFFFFC0, 1, 1, 1 } };
+        tensor_configs[1] = { GGML_TYPE_I8, { 0x7FFFFFFFFFFFFFC0, 1, 1, 1 } };
+    }
+
     if (hft == HANDCRAFTED_HEADER_BAD_N_TENSORS) {
         const uint64_t n_tensors = -1;
         helper_write(file, n_tensors);
@@ -397,7 +406,8 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         for (uint32_t i = 1; i < n_dims; ++i) {
             ne *= shape[i];
         }
-        offset += GGML_PAD(ggml_row_size(type, ne), alignment);
+
+        offset += GGML_PAD(ggml_row_size(type, ne), (uint64_t) alignment);
     }
 
     while (ftell(file) % alignment != 0) {
@@ -411,6 +421,9 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         if (hft == HANDCRAFTED_DATA_NOT_ENOUGH_DATA) {
             nbytes -= 1;
         }
+        if (hft == HANDCRAFTED_DATA_MEM_SIZE_OVERFLOW) {
+            nbytes = 32;
+        }
         for (uint64_t i = 0; i < nbytes; ++i) {
             const uint8_t random_byte = i % 256;
             helper_write(file, random_byte);
@@ -704,6 +717,7 @@ static std::pair<int, int> test_handcrafted_file(const unsigned int seed) {
         HANDCRAFTED_DATA_NOT_ENOUGH_DATA,
         HANDCRAFTED_DATA_BAD_ALIGN,
         HANDCRAFTED_DATA_INCONSISTENT_ALIGN,
+        HANDCRAFTED_DATA_MEM_SIZE_OVERFLOW,
         HANDCRAFTED_DATA_SUCCESS,
         HANDCRAFTED_DATA_CUSTOM_ALIGN,
     };

tools/server/server-http.cpp

@@ -339,6 +339,17 @@ static std::map<std::string, std::string> get_headers(const httplib::Request & r
     return headers;
 }
 
+static std::string build_query_string(const httplib::Request & req) {
+    std::string qs;
+    for (const auto & [key, value] : req.params) {
+        if (!qs.empty()) {
+            qs += '&';
+        }
+        qs += httplib::encode_query_component(key) + "=" + httplib::encode_query_component(value);
+    }
+    return qs;
+}
+
 // using unique_ptr for request to allow safe capturing in lambdas
 using server_http_req_ptr = std::unique_ptr<server_http_req>;
 
@@ -382,6 +393,7 @@ void server_http_context::get(const std::string & path, const server_http_contex
             get_params(req),
             get_headers(req),
             req.path,
+            build_query_string(req),
             req.body,
             req.is_connection_closed
         });
@@ -396,6 +408,7 @@ void server_http_context::post(const std::string & path, const server_http_conte
             get_params(req),
             get_headers(req),
             req.path,
+            build_query_string(req),
             req.body,
             req.is_connection_closed
         });

tools/server/server-http.h

@@ -36,7 +36,8 @@ using server_http_res_ptr = std::unique_ptr<server_http_res>;
 struct server_http_req {
     std::map<std::string, std::string> params; // path_params + query_params
     std::map<std::string, std::string> headers; // reserved for future use
-    std::string path; // reserved for future use
+    std::string path;
+    std::string query_string; // query parameters string (e.g. "action=save")
     std::string body;
     const std::function<bool()> & should_stop;
 

tools/server/server-models.cpp

@@ -697,11 +697,15 @@ server_http_res_ptr server_models::proxy_request(const server_http_req & req, co
         mapping[name].meta.last_used = ggml_time_ms();
     }
     SRV_INF("proxying request to model %s on port %d\n", name.c_str(), meta->port);
+    std::string proxy_path = req.path;
+    if (!req.query_string.empty()) {
+        proxy_path += '?' + req.query_string;
+    }
     auto proxy = std::make_unique<server_http_proxy>(
             method,
             CHILD_ADDR,
             meta->port,
-            req.path,
+            proxy_path,
             req.headers,
             req.body,
             req.should_stop,

tools/server/server-task.cpp

@@ -204,7 +204,8 @@ task_params server_task::params_from_json_cmpl(
     params.cache_prompt     = json_value(data,       "cache_prompt",       defaults.cache_prompt);
     params.return_tokens    = json_value(data,       "return_tokens",      false);
     params.return_progress  = json_value(data,       "return_progress",    false);
-    params.n_predict        = json_value(data,       "n_predict",          json_value(data, "max_tokens", defaults.n_predict));
+    auto max_tokens         = json_value(data,       "max_tokens",         defaults.n_predict);
+    params.n_predict        = json_value(data,       "n_predict",          json_value(data, "max_completion_tokens", max_tokens));
     params.n_indent         = json_value(data,       "n_indent",           defaults.n_indent);
     params.n_keep           = json_value(data,       "n_keep",             defaults.n_keep);
     params.n_discard        = json_value(data,       "n_discard",          defaults.n_discard);