diff --git a/common/common.cpp b/common/common.cpp
index b6638cc5..7ad73b9e 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -200,6 +200,9 @@ int32_t cpu_get_num_math() {
     return cpu_get_num_physical_cores();
 }
 
+//
+// Arg utils
+//
 common_webui common_webui_from_name(const std::string& format) {
     if (format == "none") {
         return COMMON_WEBUI_NONE;
@@ -224,6 +227,14 @@ static std::string read_file(const std::string& fname) {
     file.close();
     return content;
 }
+
+static std::string parse_device_list(const std::string& value) {
+    if (value==" " || value.find("-")!= std::string::npos) {
+        throw std::invalid_argument("no devices specified");
+    }
+    return value;
+}
+
 //
 // CLI argument parsing
 //
@@ -1066,7 +1077,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         }
         return true;
     }
-    if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--gpu-layers-draft") {
+    if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--n-gpu-layers-draft") {
         CHECK_ARG
         params.n_gpu_layers_draft = std::stoi(argv[i]);
         if (!llama_supports_gpu_offload()) {
@@ -1213,6 +1224,18 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         else { invalid_param = true; }
         return true;
     }
+    if (arg == "-dev" || arg == "--device") {
+        CHECK_ARG
+        std::string value(argv[i]);
+        params.devices = parse_device_list(value);
+        return true;
+    }
+    if (arg == "-devd" || arg == "--device-draft") {
+        CHECK_ARG
+        std::string value(argv[i]);
+        params.devices_draft = parse_device_list(value);
+        return true;
+    }
     if (arg == "-v" || arg == "--verbose") {
         params.verbosity = 1;
         return true;
@@ -2002,6 +2025,12 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
                                                                         "  - row: split rows across GPUs" });
         options.push_back({ "*",           "-ts,   --tensor-split SPLIT",
                                                                         "fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1" });
+        options.push_back({ "*",           "-dev,   --device dev1,dev2",
+                                                                         "comma-separated list of devices to use for offloading (none = don't offload)\n"
+                                                                         "Example: CUDA0,CUDA1,RPC[192.168.0.1:8080]\n" });
+        options.push_back({ "*",           "-devd,   --device-draft dev1,dev2",
+                                                                         "comma-separated list of devices to use for offloading for the draft model (none = don't offload)\n"
+                                                                         "Example: CUDA0,CUDA1,RPC[192.168.0.1:8080]\n" });
         options.push_back({ "*",           "-mg,   --main-gpu i",       "the GPU to use for the model (with split-mode = none),\n"
                                                                         "or for intermediate results and KV (with split-mode = row) (default: %d)", params.main_gpu });
     }
@@ -2575,7 +2604,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
     } else {
         model = llama_load_model_from_file(params.model.c_str(), mparams);
     }
-
+    
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
         return iparams;
@@ -2692,6 +2721,7 @@ void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lor
 
 struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
     auto mparams = llama_model_default_params();
+    mparams.devices = params.devices.c_str(); 
 
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
diff --git a/common/common.h b/common/common.h
index 905e8a25..c3d50098 100644
--- a/common/common.h
+++ b/common/common.h
@@ -126,6 +126,9 @@ struct model_paths {
 };
 
 struct gpt_params {
+    std::string devices;
+    std::string devices_draft;
+
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
     int32_t n_threads             = cpu_get_num_math();
@@ -193,6 +196,7 @@ struct gpt_params {
     std::string logits_file          = ""; // file for saving *all* logits
     std::string rpc_servers          = ""; // comma separated list of RPC servers
 
+
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
@@ -440,6 +444,7 @@ bool fs_create_directory_with_parents(const std::string & path);
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
 
+
 //
 // Model utils
 //
diff --git a/common/speculative.cpp b/common/speculative.cpp
index 5b4f8323..326a2df3 100644
--- a/common/speculative.cpp
+++ b/common/speculative.cpp
@@ -91,10 +91,10 @@ bool llama_speculative_are_compatible(
     const struct llama_vocab * vocab_dft = llama_get_model_vocab(model_dft);
 
     const bool vocab_type_tgt = llama_vocab_type(model_tgt);
-    LLAMA_LOG_INFO("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
+    LLAMA_LOG_DEBUG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
 
     const bool vocab_type_dft = llama_vocab_type(model_dft);
-    LLAMA_LOG_INFO("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
+    LLAMA_LOG_DEBUG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
 
     if (vocab_type_tgt != vocab_type_dft) {
         LLAMA_LOG_INFO("%s: draft model vocab type must match target model to use speculation but ", __func__);
@@ -203,13 +203,13 @@ std::vector<llama_token> llama_speculative_gen_draft(
         std::string text;
         text = llama_detokenize(ctx_tgt, prompt_tgt_main_model, true);
         text = replace_to_dft(spec, text);
-        LLAMA_LOG_INFO("%s: main->draft detokenized string: '%s'\n", __func__, text.c_str());
+        LLAMA_LOG_DEBUG("%s: main->draft detokenized string: '%s'\n", __func__, text.c_str());
         prompt_tgt_draft_model = llama_tokenize(ctx_dft, text, false, true);
 
         // convert id_last to draft vocab
         std::vector<llama_token> id_last_vec(1, id_last);
         text = llama_detokenize(ctx_tgt, id_last_vec);
-        LLAMA_LOG_INFO("main->draft detokenized id_last(%d): '%s'\n", id_last, text.c_str());
+        LLAMA_LOG_DEBUG("main->draft detokenized id_last(%d): '%s'\n", id_last, text.c_str());
         id_last = llama_tokenize(ctx_dft, text, false, true)[0];
     }
     // prompt_tgt's tokens will always be compatible with ctx_dft
@@ -233,8 +233,7 @@ std::vector<llama_token> llama_speculative_gen_draft(
             reuse_n = cur;
         }
     }
-
-    LLAMA_LOG_INFO("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size());
+    LLAMA_LOG_DEBUG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size());
 
     std::vector<llama_token> result;
     result.reserve(params.n_draft);
@@ -344,7 +343,7 @@ std::vector<llama_token> llama_speculative_gen_draft(
     if (!spec->vocab_dft_compatible) {
         std::string detokenized = llama_detokenize(ctx_dft, result, true);
         detokenized = replace_to_tgt(spec, detokenized);
-        LLAMA_LOG_INFO("draft->main detokenized string: '%s'\n", detokenized.c_str());
+        LLAMA_LOG_DEBUG("draft->main detokenized string: '%s'\n", detokenized.c_str());
         result = llama_tokenize(ctx_tgt, detokenized, false, true);
         if (result.size() > (size_t)params.n_draft) {
             result.resize(params.n_draft);
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index bc59480e..eb7740b6 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -1249,6 +1249,7 @@ struct server_context {
             LOG_INFO("loading draft model", {{"model", params.model_draft}});
 
             gpt_params params_dft;
+            params_dft.devices = params.devices_draft;
             params_dft.model = params.model_draft;
             params_dft.n_ctx = params.n_ctx_draft == 0 ? params.n_ctx / params.n_parallel : params.n_ctx_draft;
             params_dft.n_gpu_layers = params.n_gpu_layers_draft;
@@ -1273,7 +1274,7 @@ struct server_context {
 
             cparams_dft = llama_context_params_from_gpt_params(params_dft);
             cparams_dft.n_batch = n_ctx_dft;
-
+            
             model_draft = llama_init_dft.model;
             ctx_draft = llama_init_dft.context;
         }
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
index 3063f0b6..cb88ea1f 100644
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -71,6 +71,7 @@ int main(int argc, char ** argv) {
     ctx_tgt = llama_init_tgt.context;
 
     // load the draft model
+    params.devices = params.devices_draft;
     params.model = params.model_draft;
     params.n_gpu_layers = params.n_gpu_layers_draft;
     if (params.n_threads_draft > 0) {
diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index 05964db1..2ed282e2 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -9,6 +9,7 @@
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
+#include <string>
 #include <vector>
 #include <set>
 
@@ -528,6 +529,16 @@ GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn ini
     ggml_backend_registry_count++;
 }
 
+// Backend (reg) enumeration
+static bool striequals(const char* a, const char* b) {
+    for (; *a && *b; a++, b++) {
+        if (std::tolower(*a) != std::tolower(*b)) {
+            return false;
+        }
+    }
+    return *a == *b;
+}
+
 size_t ggml_backend_reg_get_count(void) {
     ggml_backend_registry_init();
 
@@ -539,7 +550,7 @@ size_t ggml_backend_reg_find_by_name(const char * name) {
 
     for (size_t i = 0; i < ggml_backend_registry_count; i++) {
         // TODO: case insensitive in a portable way
-        if (strcmp(ggml_backend_registry[i].name, name) == 0) {
+        if (striequals(ggml_backend_registry[i].name, name)) {
             return i;
         }
     }
diff --git a/include/llama.h b/include/llama.h
index 504f9280..046ac9c4 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -342,6 +342,9 @@ extern "C" {
     };
 
     struct llama_model_params {
+        // comma separated list of devices to use for offloading
+        const char* devices;
+
         int32_t n_gpu_layers; // number of layers to store in VRAM
         int32_t mla;          // MLA implementation to use (only applicable to DeepSeek models at this point)
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
diff --git a/src/llama-context.h b/src/llama-context.h
index 056be200..b771d459 100644
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -202,4 +202,7 @@ struct llama_context {
     struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
     struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
     struct ggml_tensor * inp_scale = nullptr; // F32 [n_tokens]
+
+    ggml_backend_t ggml_backend_by_name(const char * name);
+
 };
diff --git a/src/llama-cparams.h b/src/llama-cparams.h
index 528184c8..10a777f9 100644
--- a/src/llama-cparams.h
+++ b/src/llama-cparams.h
@@ -12,6 +12,9 @@ struct llama_cparams {
     uint32_t n_threads;       // number of threads to use for generation
     uint32_t n_threads_batch; // number of threads to use for batch processing
 
+    std::vector<std::string> devices;
+    std::vector<std::string> devices_draft;
+
     float rope_freq_base;
     float rope_freq_scale;
 
diff --git a/src/llama-impl.h b/src/llama-impl.h
index ac76e23b..36d92db8 100644
--- a/src/llama-impl.h
+++ b/src/llama-impl.h
@@ -38,6 +38,12 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
 
 #define LLAMA_LOG_INFO(...)  llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
 #define LLAMA_LOG_DEBUG(...)  llama_log_internal(GGML_LOG_LEVEL_DEBUG , __VA_ARGS__)
+#ifdef NDEBUG
+// Release mode - make LLAMA_LOG_DEBUG a no-op
+#define LLAMA_LOG_DEBUG(...) ((void)0)
+#else
+#define LLAMA_LOG_DEBUG(...) llama_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#endif
 #define LLAMA_LOG_WARN(...)  llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
 #define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
 
diff --git a/src/llama-model.h b/src/llama-model.h
index 8446be98..a26c7cb3 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -264,6 +264,7 @@ struct llama_model {
     int n_gpu_layers;
 
     std::vector<std::string> rpc_servers;
+    std::vector<int32_t> devices;
 
     // gguf metadata
     std::unordered_map<std::string, std::string> gguf_kv;
diff --git a/src/llama.cpp b/src/llama.cpp
index 2cce5384..d3ca2ace 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -154,6 +154,39 @@ static std::string trim(const std::string & str) {
     return str.substr(start, end - start);
 }
 
+
+static std::vector<std::string> llama_string_split(const std::string& str, const std::string& delimiter) {
+    std::vector<std::string> parts;
+    size_t start = 0;
+    size_t end = str.find(delimiter);
+
+    while (end != std::string::npos) {
+        parts.push_back(str.substr(start, end - start));
+        start = end + delimiter.length();
+        end = str.find(delimiter, start);
+    }
+
+    parts.push_back(str.substr(start));
+
+    return parts;
+
+}
+
+// extract ip and port from RPC[ip:port] for rpc and keep other device names
+static std::vector<std::string> extract_ip_from_rpc_device(std::vector<std::string> devices) {
+    std::vector<std::string> rpc_servers;
+    std::regex pattern("RPC\\[(.*?)\\]");
+    std::smatch matches;
+    for (auto device : devices) {
+        if (std::regex_search(device, matches, pattern)) {
+            rpc_servers.push_back(matches[1]);
+        } else {
+            rpc_servers.push_back(device);
+        }
+    }
+    return rpc_servers;
+}
+
 enum llm_chat_template {
     LLM_CHAT_TEMPLATE_CHATML,
     LLM_CHAT_TEMPLATE_LLAMA_2,
@@ -1501,6 +1534,46 @@ static void llm_prepare_mla(llama_model & model, int mla) {
     ggml_free(ctx);
 }
 
+// Backend (reg) enumeration
+static bool striequals(const char* a, const char* b) {
+    for (; *a && *b; a++, b++) {
+        if (std::tolower(*a) != std::tolower(*b)) {
+            return false;
+        }
+    }
+    return *a == *b;
+}
+
+ggml_backend_t llama_context::ggml_backend_by_name(const char* name) {
+    for (auto backend : backends) {
+        const char* backend_name = ggml_backend_name(backend);
+        if (striequals(backend_name, name)) {
+            return backend;
+        }
+    }
+    return nullptr;
+}
+
+static bool item_in_list(const std::vector<std::string>& devices, const char* name) {
+    for (auto& device : devices) {
+        if (striequals(device.c_str(), name)) {
+            return true;
+        }
+    }
+    return false;
+}
+
+static void ggml_backend_add_from_device(llama_context* ctx, ggml_backend_t backend) {
+    const char* name = ggml_backend_name(backend);
+    if (ctx->cparams.devices.size()) {
+        if (item_in_list(ctx->cparams.devices, name)) {
+            ctx->backends.push_back(backend);
+        }
+    } else {
+        ctx->backends.push_back(backend);
+    }
+}
+
 // Returns false if cancelled by progress_callback
 static bool llm_load_tensors(
         llama_model_loader & ml,
@@ -1538,13 +1611,14 @@ static bool llm_load_tensors(
 
     if (split_mode == LLAMA_SPLIT_MODE_LAYER) {
         // calculate the split points
-        int device_count = llama_get_device_count(model);
+        // int device_count = llama_get_device_count(model);
+        int device_count = model.devices.size();
         bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + device_count, [](float x) { return x == 0.0f; });
         std::vector<float> splits(device_count);
         if (all_zero) {
             // default split, by free memory
             for (int i = 0; i < device_count; ++i) {
-                splits[i] = llama_get_device_memory(model, i);
+                splits[i] = llama_get_device_memory(model, model.devices[i]);
             }
         } else {
             std::copy(tensor_split, tensor_split + device_count, splits.begin());
@@ -1564,35 +1638,35 @@ static bool llm_load_tensors(
         int act_gpu_layers = std::min(n_gpu_layers, (int)n_layer + 1);
         for (int i = i_gpu_start; i < n_layer; ++i) {
             int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(i - i_gpu_start)/act_gpu_layers) - splits.begin();
-            model.buft_layer[i] = llama_default_buffer_type_offload(model, layer_gpu);
+            model.buft_layer[i] = llama_default_buffer_type_offload(model, model.devices[layer_gpu]);
         }
         // assign the output layer
         if (n_gpu_layers > n_layer) {
             int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(act_gpu_layers - 1)/act_gpu_layers) - splits.begin();
-            model.buft_output = llama_default_buffer_type_offload(model, layer_gpu);
+            model.buft_output = llama_default_buffer_type_offload(model, model.devices[layer_gpu]);
         } else {
             model.buft_output = llama_default_buffer_type_cpu(true);
         }
     } else {
         ggml_backend_buffer_type_t split_buft;
         if (split_mode == LLAMA_SPLIT_MODE_ROW) {
-            split_buft = llama_default_buffer_type_split(model, main_gpu, tensor_split);
+            split_buft = llama_default_buffer_type_split(model, model.devices[main_gpu], tensor_split);
         } else {
             // LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_LAYER in backends where it is not supported
-            split_buft = llama_default_buffer_type_offload(model, main_gpu);
+            split_buft = llama_default_buffer_type_offload(model, model.devices[main_gpu]);
         }
         // assign the repeating layers
         for (int i = i_gpu_start; i < n_layer; ++i) {
             model.buft_layer[i] = {
                 split_buft,
-                llama_default_buffer_type_offload(model, main_gpu)
+                llama_default_buffer_type_offload(model, model.devices[main_gpu])
             };
         }
         // assign the output layer
         if (n_gpu_layers > n_layer) {
             model.buft_output = {
                 split_buft,
-                llama_default_buffer_type_offload(model, main_gpu)
+                llama_default_buffer_type_offload(model, model.devices[main_gpu])
             };
         } else {
             model.buft_output = llama_default_buffer_type_cpu(true);
@@ -3696,6 +3770,7 @@ void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
 //
 struct llama_model_params llama_model_default_params() {
     struct llama_model_params result = {
+        /*.devices                 =*/ nullptr,
         /*.n_gpu_layers                =*/ 0,
         /*.mla                         =*/ 0,
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
@@ -3859,6 +3934,17 @@ int64_t llama_time_us(void) {
     return ggml_time_us();
 }
 
+static int32_t find_device_idx(const std::string& str) {
+    std::regex pattern(R"((\d+)$)");  // Match digits at the end
+    std::smatch matches;
+    int number = -1;
+    if (std::regex_search(str, matches, pattern)) {
+        number = std::stoi(matches[1]);
+    }
+    return number;
+}
+
+
 struct llama_model * llama_load_model_from_file(
         const char * path_model,
         struct llama_model_params   params) {
@@ -3882,16 +3968,87 @@ struct llama_model * llama_load_model_from_file(
             return true;
         };
     }
-    if (params.rpc_servers != nullptr && params.rpc_servers[0] != '\0') {
-        // split the servers set them into model->rpc_servers
-        std::string servers(params.rpc_servers);
-        size_t pos = 0;
-        while ((pos = servers.find(",")) != std::string::npos) {
-            std::string server = servers.substr(0, pos);
-            model->rpc_servers.push_back(server);
-            servers.erase(0, pos + 1);
+
+    // model->devices hold device indices that are used to offload
+    // use model->devices to determine offload device
+    // if no device is specified, all device are included
+    // if device is specified, only those in the devices are included in the model->devices
+
+    std::vector<std::string> params_devices = {};
+    if (!striequals(params.devices, "")) {
+        params_devices = llama_string_split(params.devices, ",");
+        params_devices = extract_ip_from_rpc_device(params_devices);
+    }
+
+    int32_t idx = 0;
+    if (params_devices.size()) {
+        // just the number of GPU on host machine since we have not added any RPC backend
+        int dev_count = (int)llama_get_device_count(*model);
+        // list all buffer type names
+        std::vector<std::string> buffer_names = {};
+        for (int i = 0; i < dev_count; i++) {
+            ggml_backend_buffer_type_t buft = llama_default_buffer_type_offload(*model, i);
+            const char* name = ggml_backend_buft_name(buft);
+            buffer_names.push_back(std::string(name));
+        }
+
+        // add if device matches backend buffer type
+        for (auto device : params_devices) {
+            if (item_in_list(buffer_names, device.c_str())) {
+                idx = find_device_idx(device);
+                model->devices.push_back(idx);
+            } else {
+                LLAMA_LOG_ERROR("%s backend not available.\n", device.c_str());
+            }
+            
+        }
+    } else {
+        // add all backend buffer to device
+        // just the number of GPU on host machine since we have not added any RPC backend
+        int dev_count = (int)llama_get_device_count(*model);
+        for (idx = 0; idx < dev_count; idx++) {
+            model->devices.push_back(idx);
+        }
+    }
+    if (params.rpc_servers != nullptr && params.rpc_servers[0] != '\0') {
+        if (params_devices.size()) {
+            // just the number of GPU on host machine since we have not added any RPC backend
+            idx = (int)llama_get_device_count(*model); 
+            // split the servers set them into model->rpc_servers
+            std::vector <std::string> rpc_servers = llama_string_split(params.rpc_servers, ",");
+            for (auto device : params_devices) {
+                if (item_in_list(rpc_servers, device.c_str())) {
+                    model->rpc_servers.push_back(device);
+                    model->devices.push_back(idx);
+                    idx++;
+                } else {
+                    LLAMA_LOG_ERROR("%s backend not available.\n", device.c_str());
+                }
+            }
+        } else {
+                // just number of GPU on host machine since we have not added any RPC backend
+                idx = (int)llama_get_device_count(*model);
+                model->rpc_servers = llama_string_split(params.rpc_servers, ",");
+                for (auto rpc : model->rpc_servers) {
+                    model->devices.push_back(idx);
+                    idx++;
+                }
+         }
+     }
+    // no gpu used, so set layers offload to be 0
+    if (!model->devices.size()) {
+        params.n_gpu_layers = 0;
+        LLAMA_LOG_INFO("CPU: using device CPU\n");
+    } else {
+        for (auto i : model->devices) {
+            ggml_backend_buffer_type_t buft = llama_default_buffer_type_offload(*model, i);
+            const char* name = ggml_backend_buft_name(buft);
+            const char* description = name;
+            size_t description_size = llama_get_device_memory(*model, i);
+            LLAMA_LOG_INFO("%s: using device %s - %zu MiB free\n", 
+                name, description,
+                description_size / 1024 / 1024);
         }
-        model->rpc_servers.push_back(servers);
     }
     int status = llama_model_load(path_model, *model, params);
     GGML_ASSERT(status <= 0);
@@ -3948,9 +4105,18 @@ struct llama_context * llama_new_context_with_model(
 
     llama_context * ctx = new llama_context(*model);
 
+    // add devices to ctx->cparams from model
+    for (int i : model->devices) {
+        ggml_backend_buffer_type_t buft = llama_default_buffer_type_offload(*model, i);
+        const char* name = ggml_backend_buft_name(buft);
+        std::string device(name);
+        ctx->cparams.devices.push_back(device);
+    }
+
     const auto & hparams = model->hparams;
     auto       & cparams = ctx->cparams;
 
+    
     cparams.n_seq_max        = std::max(1u, params.n_seq_max);
     cparams.n_threads        = params.n_threads;
     cparams.n_threads_batch  = params.n_threads_batch;
@@ -4077,7 +4243,6 @@ struct llama_context * llama_new_context_with_model(
 
     GGML_ASSERT(hparams.n_embd_head_k % ggml_blck_size(type_k) == 0);
     GGML_ASSERT(hparams.n_embd_head_v % ggml_blck_size(type_v) == 0);
-
     if (!hparams.vocab_only) {
         // initialize backends
 #if defined(GGML_USE_METAL)
@@ -4088,7 +4253,7 @@ struct llama_context * llama_new_context_with_model(
                 llama_free(ctx);
                 return nullptr;
             }
-            ctx->backends.push_back(ctx->backend_metal);
+            ggml_backend_add_from_device(ctx,  ctx->backend_metal);
         }
 #elif defined(GGML_USE_CUDA)
         if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
@@ -4099,7 +4264,8 @@ struct llama_context * llama_new_context_with_model(
                 llama_free(ctx);
                 return nullptr;
             }
-            ctx->backends.push_back(backend);
+            ggml_backend_add_from_device(ctx, backend);
+
         } else {
             // LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU
             for (int device = 0; device < ggml_backend_cuda_get_device_count(); ++device) {
@@ -4109,7 +4275,8 @@ struct llama_context * llama_new_context_with_model(
                     llama_free(ctx);
                     return nullptr;
                 }
-                ctx->backends.push_back(backend);
+                ggml_backend_add_from_device(ctx, backend);
+
             }
         }
 #elif defined(GGML_USE_VULKAN)
@@ -4125,7 +4292,7 @@ struct llama_context * llama_new_context_with_model(
                 llama_free(ctx);
                 return nullptr;
             }
-            ctx->backends.push_back(backend);
+            ggml_backend_add_from_device(ctx, backend);
         } else {
             for (int device = 0; device < ggml_backend_vk_get_device_count(); ++device) {
                 ggml_backend_t backend = ggml_backend_vk_init(device);
@@ -4134,7 +4301,7 @@ struct llama_context * llama_new_context_with_model(
                     llama_free(ctx);
                     return nullptr;
                 }
-                ctx->backends.push_back(backend);
+                ggml_backend_add_from_device(ctx, backend);
             }
         }
 #elif defined(GGML_USE_SYCL)
@@ -4156,7 +4323,7 @@ struct llama_context * llama_new_context_with_model(
                     llama_free(ctx);
                     return nullptr;
                 }
-                ctx->backends.push_back(backend);
+                ggml_backend_add_from_device(ctx, backend);
             }
         }
 #elif defined(GGML_USE_KOMPUTE)
@@ -4167,7 +4334,7 @@ struct llama_context * llama_new_context_with_model(
                 llama_free(ctx);
                 return nullptr;
             }
-            ctx->backends.push_back(backend);
+            ggml_backend_add_from_device(ctx, backend);
         }
 #elif defined(GGML_USE_CANN)
     // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
@@ -4179,7 +4346,7 @@ struct llama_context * llama_new_context_with_model(
             llama_free(ctx);
             return nullptr;
         }
-        ctx->backends.push_back(backend);
+        ggml_backend_add_from_device(ctx, backend);
     } else {
         // LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU
         // TODO: currently, CANN can't use multi-gpus, just leave code here for further cann version.
@@ -4190,7 +4357,7 @@ struct llama_context * llama_new_context_with_model(
                 llama_free(ctx);
                 return nullptr;
             }
-            ctx->backends.push_back(backend);
+            ggml_backend_add_from_device(ctx, backend);
         }
     }
 #endif
@@ -4200,7 +4367,7 @@ struct llama_context * llama_new_context_with_model(
         if (ctx->backend_blas == nullptr) {
             LLAMA_LOG_WARN("%s: failed to initialize BLAS backend\n", __func__);
         } else {
-            ctx->backends.push_back(ctx->backend_blas);
+            ggml_backend_add_from_device(ctx, ctx->backend_blas);
         }
 #endif
 
@@ -4213,10 +4380,23 @@ struct llama_context * llama_new_context_with_model(
                     llama_free(ctx);
                     return nullptr;
                 }
-                ctx->backends.push_back(backend);
+                ggml_backend_add_from_device(ctx, backend);
             }
         }
 #endif
+        if (ctx->cparams.devices.size()) {
+            // reorder the backend from devices params
+            std::vector<ggml_backend_t> backends = {};
+            std::vector<const char*> device_list = {};
+            for (auto device : ctx->cparams.devices) {
+                ggml_backend_t backend = ctx->ggml_backend_by_name(device.c_str());
+                if (backend) {
+                    backends.push_back(backend);
+                }
+            }
+            ctx->backends = std::move(backends);
+        }
+        
         ctx->backend_cpu = ggml_backend_cpu_init();
         if (ctx->backend_cpu == nullptr) {
             LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__);