diff --git a/examples/llava/README.md b/examples/llava/README.md
index 25ea96715..35e6d9e5d 100644
--- a/examples/llava/README.md
+++ b/examples/llava/README.md
@@ -63,13 +63,12 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director
 ```console
 git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
 ```
-2) Backup your pth/safetensor model files as llava-surgery modifies them
-3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
+2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
 ```console
 python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
 ```
 - you will find a llava.projector and a llava.clip file in your model directory
-4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
+3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
 ```console
 mkdir vit
 cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
@@ -77,18 +76,18 @@ cp ../llava-v1.6-vicuna-7b/llava.projector vit/
 curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
 ```
 
-5) Create the visual gguf model:
+4) Create the visual gguf model:
 ```console
 python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
 ```
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
 
-6) Then convert the model to gguf format:
+5) Then convert the model to gguf format:
 ```console
-python ./convert.py ../llava-v1.6-vicuna-7b/
+python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
 ```
 
-7) And finally we can run the llava-cli using the 1.6 model version:
+6) And finally we can run the llava-cli using the 1.6 model version:
 ```console
 ./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
 ```
diff --git a/examples/llava/llava-surgery-v2.py b/examples/llava/llava-surgery-v2.py
index 5bc5bc513..eb56d6988 100644
--- a/examples/llava/llava-surgery-v2.py
+++ b/examples/llava/llava-surgery-v2.py
@@ -65,9 +65,7 @@ def clean_vision_tower_from_checkpoint(checkpoint_path):
         for name in clip_tensors:
             del checkpoint[name]
 
-        # Save the updated checkpoint
         checkpoint_path = checkpoint_path
-        save_model(checkpoint, checkpoint_path, file_type)
         return True
     return False
 
@@ -152,16 +150,6 @@ for name in first_mm_tensors:
 if len(projector) > 0:
     save_model(projector, f"{args.model}/llava.projector", 'pytorch')
 
-for name in mm_tensors:
-    del last_checkpoint[name]
-for name in first_mm_tensors:
-    del first_checkpoint[name]
-
-if len(mm_tensors) > 0:
-    save_model(last_checkpoint, projector_checkpoint_path, file_type)
-if len(first_mm_tensors) > 0:
-    save_model(first_checkpoint, newline_checkpoint_path, file_type)
-
 print("Done!")
 print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
 print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index dae3703fc..8c49a626e 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -335,6 +335,8 @@ int main(int argc, char ** argv) {
     // number of tokens to keep when resetting context
     if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) {
         params.n_keep = (int)embd_inp.size();
+    } else {
+        params.n_keep += add_bos; // always keep the BOS token
     }
 
     // prefix & suffix for instruct mode
@@ -384,8 +386,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        if (params.n_keep > 0) {
-        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+        if (params.n_keep > add_bos) {
+            LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
                 LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
             }
@@ -541,14 +543,14 @@ int main(int argc, char ** argv) {
                         break;
                     }
 
-                    const int n_left    = n_past - params.n_keep - 1;
+                    const int n_left    = n_past - params.n_keep;
                     const int n_discard = n_left/2;
 
                     LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                             n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 
diff --git a/examples/server/README.md b/examples/server/README.md
index f6b9c7402..4b24ee5dc 100644
--- a/examples/server/README.md
+++ b/examples/server/README.md
@@ -41,6 +41,7 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
 - `-n, --n-predict`: Set the maximum tokens to predict (default: -1)
 - `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
+- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
 
 ## Build
 
@@ -140,6 +141,8 @@ node index.js
   - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available.
   - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slot are currently available.
 
+  If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
+
 - **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index 517e0257f..635a52603 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -1395,6 +1395,46 @@ struct llama_server_context
             case TASK_TYPE_NEXT_RESPONSE: {
                 // do nothing
             } break;
+            case TASK_TYPE_SLOTS_DATA: {
+                json slots_data        = json::array();
+                int n_idle_slots       = 0;
+                int n_processing_slots = 0;
+
+                for (llama_client_slot &slot: slots) {
+                    if (slot.available()) {
+                        n_idle_slots++;
+                    } else {
+                        n_processing_slots++;
+                    }
+                    json slot_data = get_formated_generation(slot);
+                    slot_data["id"] = slot.id;
+                    slot_data["task_id"] = slot.task_id;
+                    slot_data["state"] = slot.state;
+                    slot_data["prompt"] = slot.prompt;
+                    slot_data["next_token"] = {
+                            {"has_next_token", slot.has_next_token},
+                            {"n_remain", slot.n_remaining},
+                            {"num_tokens_predicted", slot.n_decoded},
+                            {"stopped_eos", slot.stopped_eos},
+                            {"stopped_word", slot.stopped_word},
+                            {"stopped_limit", slot.stopped_limit},
+                            {"stopping_word", slot.stopping_word},
+                    };
+                    slots_data.push_back(slot_data);
+                }
+                LOG_TEE("task %i - slots data: idle=%i processing=%i\n", task.id, n_idle_slots, n_processing_slots);
+                task_result res;
+                res.id = task.id;
+                res.multitask_id = task.multitask_id;
+                res.stop = true;
+                res.error = false;
+                res.result_json = {
+                        { "idle",       n_idle_slots       },
+                        { "processing", n_processing_slots },
+                        { "slots",      slots_data         }
+                };
+                queue_results.send(res);
+            } break;
         }
     }
 
@@ -1448,14 +1488,15 @@ struct llama_server_context
                 if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx)
                 {
                     // Shift context
-                    const int n_left    = system_tokens.size() + slot.n_past - slot.params.n_keep - 1;
+                    const int n_keep    = slot.params.n_keep + add_bos_token;
+                    const int n_left    = system_tokens.size() + slot.n_past - n_keep;
                     const int n_discard = n_left / 2;
 
-                    LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard);
-                    llama_kv_cache_seq_rm   (ctx, slot.id, slot.params.n_keep + 1            , slot.params.n_keep + n_discard + 1);
-                    llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, system_tokens.size() + slot.n_past, -n_discard);
+                    LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, n_keep, n_left, n_discard);
+                    llama_kv_cache_seq_rm   (ctx, slot.id, n_keep            , n_keep + n_discard);
+                    llama_kv_cache_seq_shift(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
 
-                    for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++)
+                    for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++)
                     {
                         slot.cache_tokens[i - n_discard] = slot.cache_tokens[i];
                     }
@@ -1468,7 +1509,7 @@ struct llama_server_context
 
                     LOG_VERBOSE("context shift", {
                         { "n_ctx", n_ctx },
-                        { "n_keep", params.n_keep },
+                        { "n_keep", n_keep },
                         { "n_left", n_left },
                     });
                 }
@@ -2558,34 +2599,38 @@ int main(int argc, char **argv)
         server_state current_state = state.load();
         switch(current_state) {
             case SERVER_STATE_READY: {
-                int available_slots  = 0;
-                int processing_slots = 0;
-                for (llama_client_slot &slot: llama.slots) {
-                    if (slot.available()) {
-                        available_slots++;
-                    } else {
-                        processing_slots++;
-                    }
+                // request slots data using task queue
+                task_server task;
+                task.id   = llama.queue_tasks.get_new_id();
+                task.type = TASK_TYPE_SLOTS_DATA;
+                task.target_id = -1;
+
+                llama.queue_results.add_waiting_task_id(task.id);
+                llama.queue_tasks.post(task);
+
+                // get the result
+                task_result result = llama.queue_results.recv(task.id);
+                llama.queue_results.remove_waiting_task_id(task.id);
+
+                int n_idle_slots       = result.result_json["idle"];
+                int n_processing_slots = result.result_json["processing"];
+
+                json health = {
+                        {"status",           "ok"},
+                        {"slots_idle",       n_idle_slots},
+                        {"slots_processing", n_processing_slots}};
+                res.status = 200; // HTTP OK
+                if (sparams.slots_endpoint && req.has_param("include_slots")) {
+                    health["slots"] = result.result_json["slots"];
                 }
-                if (available_slots > 0) {
-                    json health = {
-                            {"status",           "ok"},
-                            {"slots_idle",       available_slots},
-                            {"slots_processing", processing_slots}};
-                    res.set_content(health.dump(), "application/json");
-                    res.status = 200; // HTTP OK
-                } else {
-                    json health = {
-                            {"status",           "no slot available"},
-                            {"slots_idle",       available_slots},
-                            {"slots_processing", processing_slots}};
-                    res.set_content(health.dump(), "application/json");
+
+                if (n_idle_slots == 0) {
+                    health["status"] = "no slot available";
                     if (req.has_param("fail_on_no_slot")) {
                         res.status = 503; // HTTP Service Unavailable
-                    } else {
-                        res.status = 200; // HTTP OK
                     }
                 }
+                res.set_content(health.dump(), "application/json");
                 break;
             }
             case SERVER_STATE_LOADING_MODEL:
@@ -2601,26 +2646,20 @@ int main(int argc, char **argv)
 
     if (sparams.slots_endpoint) {
         svr.Get("/slots", [&](const httplib::Request&, httplib::Response& res) {
-            json slots;
-            for (llama_client_slot & slot : llama.slots) {
-                json slot_data = llama.get_formated_generation(slot);
-                slot_data["id"] = slot.id;
-                slot_data["task_id"] = slot.task_id;
-                slot_data["state"] = slot.state;
-                slot_data["prompt"] = slot.prompt;
-                slot_data["next_token"] = {
-                        {"has_next_token", slot.has_next_token},
-                        {"n_remain", slot.n_remaining},
-                        {"num_tokens_predicted", slot.n_decoded},
-                        {"stopped_eos", slot.stopped_eos},
-                        {"stopped_word", slot.stopped_word},
-                        {"stopped_limit", slot.stopped_limit},
-                        {"stopping_word", slot.stopping_word},
-                };
+            // request slots data using task queue
+            task_server task;
+            task.id = llama.queue_tasks.get_new_id();
+            task.type = TASK_TYPE_SLOTS_DATA;
+            task.target_id = -1;
 
-                slots.push_back(slot_data);
-            }
-            res.set_content(slots.dump(), "application/json");
+            llama.queue_results.add_waiting_task_id(task.id);
+            llama.queue_tasks.post(task);
+
+            // get the result
+            task_result result = llama.queue_results.recv(task.id);
+            llama.queue_results.remove_waiting_task_id(task.id);
+
+            res.set_content(result.result_json["slots"].dump(), "application/json");
             res.status = 200; // HTTP OK
         });
     }
diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp
index e954fb0ef..88545eb69 100644
--- a/examples/server/utils.hpp
+++ b/examples/server/utils.hpp
@@ -49,7 +49,8 @@ enum server_state {
 enum task_type {
     TASK_TYPE_COMPLETION,
     TASK_TYPE_CANCEL,
-    TASK_TYPE_NEXT_RESPONSE
+    TASK_TYPE_NEXT_RESPONSE,
+    TASK_TYPE_SLOTS_DATA
 };
 
 struct task_server {
diff --git a/ggml.c b/ggml.c
index 7df3933b6..670068265 100644
--- a/ggml.c
+++ b/ggml.c
@@ -5645,7 +5645,9 @@ struct ggml_tensor * ggml_conv_2d(
                 ggml_reshape_2d(ctx, im2col, im2col->ne[0],  im2col->ne[3] * im2col->ne[2] * im2col->ne[1]), // [N, OH, OW, IC * KH * KW] => [N*OH*OW, IC * KH * KW]
                 ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1] * a->ne[2]),  a->ne[3]));                       // [OC，IC, KH, KW] => [OC, IC * KH * KW]
 
-    result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], a->ne[3], im2col->ne[3]); // [N, OC, OH, OW]
+    result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], im2col->ne[3], a->ne[3]); // [OC, N, OH, OW]
+    result = ggml_cont(ctx, ggml_permute(ctx, result, 0, 1, 3, 2)); // [N, OC, OH, OW]
+
 
     return result;
 }
@@ -6651,8 +6653,10 @@ void ggml_set_param(
 
 static void ggml_compute_forward_dup_same_cont(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
     GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
     GGML_ASSERT(src0->type == dst->type);
@@ -6683,8 +6687,10 @@ static void ggml_compute_forward_dup_same_cont(
 }
 static void ggml_compute_forward_dup_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -6697,7 +6703,7 @@ static void ggml_compute_forward_dup_f16(
     const int nth = params->nth; // number of threads
 
     if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) {
-        ggml_compute_forward_dup_same_cont(params, src0, dst);
+        ggml_compute_forward_dup_same_cont(params, dst);
         return;
     }
 
@@ -6954,8 +6960,10 @@ static void ggml_compute_forward_dup_f16(
 
 static void ggml_compute_forward_dup_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -6968,7 +6976,7 @@ static void ggml_compute_forward_dup_f32(
     const int nth = params->nth; // number of threads
 
     if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) {
-        ggml_compute_forward_dup_same_cont(params, src0, dst);
+        ggml_compute_forward_dup_same_cont(params, dst);
         return;
     }
 
@@ -7204,8 +7212,10 @@ static void ggml_compute_forward_dup_f32(
 // A simplified version of ggml_compute_forward_dup that doesn't do float upcasting, and just plain old memcpy.
 static void ggml_compute_forward_dup_bytes(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
     GGML_ASSERT(src0->type == dst->type);
 
@@ -7214,7 +7224,7 @@ static void ggml_compute_forward_dup_bytes(
     }
 
     if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) {
-        ggml_compute_forward_dup_same_cont(params, src0, dst);
+        ggml_compute_forward_dup_same_cont(params, dst);
         return;
     }
 
@@ -7353,21 +7363,23 @@ static void ggml_compute_forward_dup_bytes(
 
 static void ggml_compute_forward_dup(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     if (src0->type == dst->type) {
-        ggml_compute_forward_dup_bytes(params, src0, dst);
+        ggml_compute_forward_dup_bytes(params, dst);
         return;
     }
 
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_dup_f16(params, src0, dst);
+                ggml_compute_forward_dup_f16(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_dup_f32(params, src0, dst);
+                ggml_compute_forward_dup_f32(params, dst);
             } break;
         default:
             {
@@ -7380,9 +7392,11 @@ static void ggml_compute_forward_dup(
 
 static void ggml_compute_forward_add_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -7468,9 +7482,11 @@ static void ggml_compute_forward_add_f32(
 
 static void ggml_compute_forward_add_f16_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -7545,9 +7561,11 @@ static void ggml_compute_forward_add_f16_f32(
 
 static void ggml_compute_forward_add_f16_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -7599,9 +7617,11 @@ static void ggml_compute_forward_add_f16_f16(
 
 static void ggml_compute_forward_add_q_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -7677,14 +7697,16 @@ static void ggml_compute_forward_add_q_f32(
 
 static void ggml_compute_forward_add(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
                 if (src1->type == GGML_TYPE_F32) {
-                    ggml_compute_forward_add_f32(params, src0, src1, dst);
+                    ggml_compute_forward_add_f32(params, dst);
                 }
                 else {
                     GGML_ASSERT(false);
@@ -7693,10 +7715,10 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_F16:
             {
                 if (src1->type == GGML_TYPE_F16) {
-                    ggml_compute_forward_add_f16_f16(params, src0, src1, dst);
+                    ggml_compute_forward_add_f16_f16(params, dst);
                 }
                 else if (src1->type == GGML_TYPE_F32) {
-                    ggml_compute_forward_add_f16_f32(params, src0, src1, dst);
+                    ggml_compute_forward_add_f16_f32(params, dst);
                 }
                 else {
                     GGML_ASSERT(false);
@@ -7718,7 +7740,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
             {
-                ggml_compute_forward_add_q_f32(params, src0, src1, dst);
+                ggml_compute_forward_add_q_f32(params, dst);
             } break;
         default:
             {
@@ -7731,9 +7753,11 @@ static void ggml_compute_forward_add(
 
 static void ggml_compute_forward_add1_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_scalar(src1));
 
@@ -7783,9 +7807,11 @@ static void ggml_compute_forward_add1_f32(
 
 static void ggml_compute_forward_add1_f16_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_scalar(src1));
 
@@ -7833,9 +7859,11 @@ static void ggml_compute_forward_add1_f16_f32(
 
 static void ggml_compute_forward_add1_f16_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_scalar(src1));
 
@@ -7883,9 +7911,11 @@ static void ggml_compute_forward_add1_f16_f16(
 
 static void ggml_compute_forward_add1_q_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_scalar(src1));
 
@@ -7950,21 +7980,23 @@ static void ggml_compute_forward_add1_q_f32(
 
 static void ggml_compute_forward_add1(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_add1_f32(params, src0, src1, dst);
+                ggml_compute_forward_add1_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
             {
                 if (src1->type == GGML_TYPE_F16) {
-                    ggml_compute_forward_add1_f16_f16(params, src0, src1, dst);
+                    ggml_compute_forward_add1_f16_f16(params, dst);
                 }
                 else if (src1->type == GGML_TYPE_F32) {
-                    ggml_compute_forward_add1_f16_f32(params, src0, src1, dst);
+                    ggml_compute_forward_add1_f16_f32(params, dst);
                 }
                 else {
                     GGML_ASSERT(false);
@@ -7987,7 +8019,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
             {
-                ggml_compute_forward_add1_q_f32(params, src0, src1, dst);
+                ggml_compute_forward_add1_q_f32(params, dst);
             } break;
         default:
             {
@@ -8000,9 +8032,11 @@ static void ggml_compute_forward_add1(
 
 static void ggml_compute_forward_acc_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
 
@@ -8082,14 +8116,14 @@ static void ggml_compute_forward_acc_f32(
 
 static void ggml_compute_forward_acc(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_acc_f32(params, src0, src1, dst);
+                ggml_compute_forward_acc_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_Q4_0:
@@ -8119,9 +8153,11 @@ static void ggml_compute_forward_acc(
 
 static void ggml_compute_forward_sub_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
@@ -8179,13 +8215,14 @@ static void ggml_compute_forward_sub_f32(
 
 static void ggml_compute_forward_sub(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sub_f32(params, src0, src1, dst);
+                ggml_compute_forward_sub_f32(params, dst);
             } break;
         default:
             {
@@ -8198,9 +8235,11 @@ static void ggml_compute_forward_sub(
 
 static void ggml_compute_forward_mul_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -8281,15 +8320,17 @@ static void ggml_compute_forward_mul_f32(
 
 static void ggml_compute_forward_mul(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src1->type == GGML_TYPE_F32 && "only f32 src1 supported for now");
 
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_mul_f32(params, src0, src1, dst);
+                ggml_compute_forward_mul_f32(params, dst);
             } break;
         default:
             {
@@ -8302,9 +8343,11 @@ static void ggml_compute_forward_mul(
 
 static void ggml_compute_forward_div_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -8375,13 +8418,14 @@ static void ggml_compute_forward_div_f32(
 
 static void ggml_compute_forward_div(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_div_f32(params, src0, src1, dst);
+                ggml_compute_forward_div_f32(params, dst);
             } break;
         default:
             {
@@ -8394,8 +8438,10 @@ static void ggml_compute_forward_div(
 
 static void ggml_compute_forward_sqr_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -8418,12 +8464,14 @@ static void ggml_compute_forward_sqr_f32(
 
 static void ggml_compute_forward_sqr(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sqr_f32(params, src0, dst);
+                ggml_compute_forward_sqr_f32(params, dst);
             } break;
         default:
             {
@@ -8436,8 +8484,10 @@ static void ggml_compute_forward_sqr(
 
 static void ggml_compute_forward_sqrt_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -8460,12 +8510,14 @@ static void ggml_compute_forward_sqrt_f32(
 
 static void ggml_compute_forward_sqrt(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sqrt_f32(params, src0, dst);
+                ggml_compute_forward_sqrt_f32(params, dst);
             } break;
         default:
             {
@@ -8478,8 +8530,10 @@ static void ggml_compute_forward_sqrt(
 
 static void ggml_compute_forward_log_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(params->ith == 0);
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
@@ -8502,12 +8556,14 @@ static void ggml_compute_forward_log_f32(
 
 static void ggml_compute_forward_log(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_log_f32(params, src0, dst);
+                ggml_compute_forward_log_f32(params, dst);
             } break;
         default:
             {
@@ -8520,8 +8576,10 @@ static void ggml_compute_forward_log(
 
 static void ggml_compute_forward_sum_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_is_scalar(dst));
 
@@ -8553,8 +8611,10 @@ static void ggml_compute_forward_sum_f32(
 
 static void ggml_compute_forward_sum_f16(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
           struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_is_scalar(dst));
 
@@ -8585,16 +8645,18 @@ static void ggml_compute_forward_sum_f16(
 
 static void ggml_compute_forward_sum(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sum_f32(params, src0, dst);
+                ggml_compute_forward_sum_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_sum_f16(params, src0, dst);
+                ggml_compute_forward_sum_f16(params, dst);
             } break;
         default:
             {
@@ -8607,8 +8669,10 @@ static void ggml_compute_forward_sum(
 
 static void ggml_compute_forward_sum_rows_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -8640,12 +8704,14 @@ static void ggml_compute_forward_sum_rows_f32(
 
 static void ggml_compute_forward_sum_rows(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sum_rows_f32(params, src0, dst);
+                ggml_compute_forward_sum_rows_f32(params, dst);
             } break;
         default:
             {
@@ -8658,8 +8724,10 @@ static void ggml_compute_forward_sum_rows(
 
 static void ggml_compute_forward_mean_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -8695,12 +8763,14 @@ static void ggml_compute_forward_mean_f32(
 
 static void ggml_compute_forward_mean(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_mean_f32(params, src0, dst);
+                ggml_compute_forward_mean_f32(params, dst);
             } break;
         default:
             {
@@ -8713,8 +8783,10 @@ static void ggml_compute_forward_mean(
 
 static void ggml_compute_forward_argmax_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -8741,12 +8813,14 @@ static void ggml_compute_forward_argmax_f32(
 
 static void ggml_compute_forward_argmax(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_argmax_f32(params, src0, dst);
+                ggml_compute_forward_argmax_f32(params, dst);
             } break;
         default:
             {
@@ -8759,8 +8833,10 @@ static void ggml_compute_forward_argmax(
 
 static void ggml_compute_forward_repeat_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(params->ith == 0);
     GGML_ASSERT(ggml_can_repeat(src0, dst));
 
@@ -8802,8 +8878,10 @@ static void ggml_compute_forward_repeat_f32(
 
 static void ggml_compute_forward_repeat_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(params->ith == 0);
     GGML_ASSERT(ggml_can_repeat(src0, dst));
 
@@ -8848,18 +8926,20 @@ static void ggml_compute_forward_repeat_f16(
 
 static void ggml_compute_forward_repeat(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
         case GGML_TYPE_I16:
             {
-                ggml_compute_forward_repeat_f16(params, src0, dst);
+                ggml_compute_forward_repeat_f16(params, dst);
             } break;
         case GGML_TYPE_F32:
         case GGML_TYPE_I32:
             {
-                ggml_compute_forward_repeat_f32(params, src0, dst);
+                ggml_compute_forward_repeat_f32(params, dst);
             } break;
         default:
             {
@@ -8872,8 +8952,10 @@ static void ggml_compute_forward_repeat(
 
 static void ggml_compute_forward_repeat_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(params->ith == 0);
     GGML_ASSERT(ggml_can_repeat(dst, src0));
 
@@ -8929,12 +9011,14 @@ static void ggml_compute_forward_repeat_back_f32(
 
 static void ggml_compute_forward_repeat_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_repeat_back_f32(params, src0, dst);
+                ggml_compute_forward_repeat_back_f32(params, dst);
             } break;
         default:
             {
@@ -8947,10 +9031,11 @@ static void ggml_compute_forward_repeat_back(
 
 static void ggml_compute_forward_concat_f32(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
-    const struct ggml_tensor * src1,
     struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -8995,14 +9080,15 @@ static void ggml_compute_forward_concat_f32(
 
 static void ggml_compute_forward_concat(
     const struct ggml_compute_params* params,
-    const struct ggml_tensor* src0,
-    const struct ggml_tensor* src1,
     struct ggml_tensor* dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_I32:
             {
-                ggml_compute_forward_concat_f32(params, src0, src1, dst);
+                ggml_compute_forward_concat_f32(params, dst);
             } break;
         default:
             {
@@ -9015,8 +9101,10 @@ static void ggml_compute_forward_concat(
 
 static void ggml_compute_forward_abs_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -9039,12 +9127,14 @@ static void ggml_compute_forward_abs_f32(
 
 static void ggml_compute_forward_abs(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_abs_f32(params, src0, dst);
+                ggml_compute_forward_abs_f32(params, dst);
             } break;
         default:
             {
@@ -9057,8 +9147,10 @@ static void ggml_compute_forward_abs(
 
 static void ggml_compute_forward_sgn_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -9081,12 +9173,14 @@ static void ggml_compute_forward_sgn_f32(
 
 static void ggml_compute_forward_sgn(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sgn_f32(params, src0, dst);
+                ggml_compute_forward_sgn_f32(params, dst);
             } break;
         default:
             {
@@ -9099,8 +9193,10 @@ static void ggml_compute_forward_sgn(
 
 static void ggml_compute_forward_neg_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -9123,12 +9219,14 @@ static void ggml_compute_forward_neg_f32(
 
 static void ggml_compute_forward_neg(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_neg_f32(params, src0, dst);
+                ggml_compute_forward_neg_f32(params, dst);
             } break;
         default:
             {
@@ -9141,8 +9239,10 @@ static void ggml_compute_forward_neg(
 
 static void ggml_compute_forward_step_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -9165,12 +9265,14 @@ static void ggml_compute_forward_step_f32(
 
 static void ggml_compute_forward_step(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_step_f32(params, src0, dst);
+                ggml_compute_forward_step_f32(params, dst);
             } break;
         default:
             {
@@ -9183,8 +9285,10 @@ static void ggml_compute_forward_step(
 
 static void ggml_compute_forward_tanh_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -9207,12 +9311,14 @@ static void ggml_compute_forward_tanh_f32(
 
 static void ggml_compute_forward_tanh(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_tanh_f32(params, src0, dst);
+                ggml_compute_forward_tanh_f32(params, dst);
             } break;
         default:
             {
@@ -9225,8 +9331,10 @@ static void ggml_compute_forward_tanh(
 
 static void ggml_compute_forward_elu_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -9249,12 +9357,14 @@ static void ggml_compute_forward_elu_f32(
 
 static void ggml_compute_forward_elu(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_elu_f32(params, src0, dst);
+                ggml_compute_forward_elu_f32(params, dst);
             } break;
         default:
             {
@@ -9267,8 +9377,10 @@ static void ggml_compute_forward_elu(
 
 static void ggml_compute_forward_relu_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -9291,12 +9403,14 @@ static void ggml_compute_forward_relu_f32(
 
 static void ggml_compute_forward_relu(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_relu_f32(params, src0, dst);
+                ggml_compute_forward_relu_f32(params, dst);
             } break;
         default:
             {
@@ -9309,8 +9423,10 @@ static void ggml_compute_forward_relu(
 
 static void ggml_compute_forward_gelu_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
     GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
@@ -9350,12 +9466,14 @@ static void ggml_compute_forward_gelu_f32(
 
 static void ggml_compute_forward_gelu(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_gelu_f32(params, src0, dst);
+                ggml_compute_forward_gelu_f32(params, dst);
             } break;
         default:
             {
@@ -9368,8 +9486,10 @@ static void ggml_compute_forward_gelu(
 
 static void ggml_compute_forward_gelu_quick_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
     GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
@@ -9409,12 +9529,14 @@ static void ggml_compute_forward_gelu_quick_f32(
 
 static void ggml_compute_forward_gelu_quick(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_gelu_quick_f32(params, src0, dst);
+                ggml_compute_forward_gelu_quick_f32(params, dst);
             } break;
         default:
             {
@@ -9427,8 +9549,10 @@ static void ggml_compute_forward_gelu_quick(
 
 static void ggml_compute_forward_silu_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
     GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
@@ -9468,12 +9592,14 @@ static void ggml_compute_forward_silu_f32(
 
 static void ggml_compute_forward_silu(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_silu_f32(params, src0, dst);
+                ggml_compute_forward_silu_f32(params, dst);
             } break;
         default:
             {
@@ -9485,8 +9611,10 @@ static void ggml_compute_forward_silu(
 
 static void ggml_compute_forward_leaky_relu_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -9512,12 +9640,14 @@ static void ggml_compute_forward_leaky_relu_f32(
 
 static void ggml_compute_forward_leaky_relu(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_leaky_relu_f32(params, src0, dst);
+                ggml_compute_forward_leaky_relu_f32(params, dst);
             } break;
         default:
             {
@@ -9530,9 +9660,11 @@ static void ggml_compute_forward_leaky_relu(
 
 static void ggml_compute_forward_silu_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * grad,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * grad = dst->src[1];
+
     GGML_ASSERT(ggml_is_contiguous_except_dim_1(grad));
     GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
     GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
@@ -9575,13 +9707,14 @@ static void ggml_compute_forward_silu_back_f32(
 
 static void ggml_compute_forward_silu_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * grad,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_silu_back_f32(params, src0, grad, dst);
+                ggml_compute_forward_silu_back_f32(params, dst);
             } break;
         default:
             {
@@ -9593,8 +9726,10 @@ static void ggml_compute_forward_silu_back(
 
 static void ggml_compute_forward_hardswish_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -9616,12 +9751,14 @@ static void ggml_compute_forward_hardswish_f32(
 }
 static void ggml_compute_forward_hardswish(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_hardswish_f32(params, src0, dst);
+                ggml_compute_forward_hardswish_f32(params, dst);
             } break;
         default:
             {
@@ -9632,8 +9769,10 @@ static void ggml_compute_forward_hardswish(
 
 static void ggml_compute_forward_hardsigmoid_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, dst));
 
@@ -9656,12 +9795,14 @@ static void ggml_compute_forward_hardsigmoid_f32(
 
 static void ggml_compute_forward_hardsigmoid(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_hardsigmoid_f32(params, src0, dst);
+                ggml_compute_forward_hardsigmoid_f32(params, dst);
             } break;
         default:
             {
@@ -9675,8 +9816,10 @@ static void ggml_compute_forward_hardsigmoid(
 
 static void ggml_compute_forward_norm_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -9728,12 +9871,14 @@ static void ggml_compute_forward_norm_f32(
 
 static void ggml_compute_forward_norm(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_norm_f32(params, src0, dst);
+                ggml_compute_forward_norm_f32(params, dst);
             } break;
         default:
             {
@@ -9746,8 +9891,10 @@ static void ggml_compute_forward_norm(
 
 static void ggml_compute_forward_rms_norm_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -9796,12 +9943,14 @@ static void ggml_compute_forward_rms_norm_f32(
 
 static void ggml_compute_forward_rms_norm(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rms_norm_f32(params, src0, dst);
+                ggml_compute_forward_rms_norm_f32(params, dst);
             } break;
         default:
             {
@@ -9812,9 +9961,11 @@ static void ggml_compute_forward_rms_norm(
 
 static void ggml_compute_forward_rms_norm_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_are_same_shape(src0, src1));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -9969,13 +10120,14 @@ static void ggml_compute_forward_rms_norm_back_f32(
 
 static void ggml_compute_forward_rms_norm_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rms_norm_back_f32(params, src0, src1, dst);
+                ggml_compute_forward_rms_norm_back_f32(params, dst);
             } break;
         default:
             {
@@ -9988,8 +10140,10 @@ static void ggml_compute_forward_rms_norm_back(
 
 static void ggml_compute_forward_group_norm_f32(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
     struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -10060,12 +10214,14 @@ static void ggml_compute_forward_group_norm_f32(
 
 static void ggml_compute_forward_group_norm(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
     struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_group_norm_f32(params, src0, dst);
+                ggml_compute_forward_group_norm_f32(params, dst);
             } break;
         default:
             {
@@ -10111,9 +10267,11 @@ static bool ggml_compute_forward_mul_mat_use_blas(struct ggml_tensor * dst) {
 
 static void ggml_compute_forward_mul_mat(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
@@ -10358,10 +10516,11 @@ static void ggml_compute_forward_mul_mat(
 
 static void ggml_compute_forward_mul_mat_id(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * ids,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * ids = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     const struct ggml_tensor * src0 = dst->src[2]; // only for GGML_TENSOR_BINARY_OP_LOCALS
 
     GGML_TENSOR_BINARY_OP_LOCALS
@@ -10552,9 +10711,11 @@ static void ggml_compute_forward_mul_mat_id(
 
 static void ggml_compute_forward_out_prod_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     // int64_t t0 = ggml_perf_time_us();
     // UNUSED(t0);
 
@@ -10744,9 +10905,11 @@ static void ggml_compute_forward_out_prod_f32(
 
 static void ggml_compute_forward_out_prod_q_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     // int64_t t0 = ggml_perf_time_us();
     // UNUSED(t0);
 
@@ -10857,9 +11020,10 @@ static void ggml_compute_forward_out_prod_q_f32(
 
 static void ggml_compute_forward_out_prod(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
@@ -10877,16 +11041,16 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
             {
-                ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst);
+                ggml_compute_forward_out_prod_q_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
             {
                 GGML_ASSERT(false); // todo
-                // ggml_compute_forward_out_prod_f16_f32(params, src0, src1, dst);
+                // ggml_compute_forward_out_prod_f16_f32(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_out_prod_f32(params, src0, src1, dst);
+                ggml_compute_forward_out_prod_f32(params, dst);
             } break;
         default:
             {
@@ -10899,8 +11063,10 @@ static void ggml_compute_forward_out_prod(
 
 static void ggml_compute_forward_scale_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(dst));
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
@@ -10941,12 +11107,14 @@ static void ggml_compute_forward_scale_f32(
 
 static void ggml_compute_forward_scale(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_scale_f32(params, src0, dst);
+                ggml_compute_forward_scale_f32(params, dst);
             } break;
         default:
             {
@@ -10959,9 +11127,11 @@ static void ggml_compute_forward_scale(
 
 static void ggml_compute_forward_set_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
 
@@ -11032,14 +11202,14 @@ static void ggml_compute_forward_set_f32(
 
 static void ggml_compute_forward_set(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_set_f32(params, src0, src1, dst);
+                ggml_compute_forward_set_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_Q4_0:
@@ -11069,29 +11239,25 @@ static void ggml_compute_forward_set(
 
 static void ggml_compute_forward_cpy(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    ggml_compute_forward_dup(params, src0, dst);
+    ggml_compute_forward_dup(params, dst);
 }
 
 // ggml_compute_forward_cont
 
 static void ggml_compute_forward_cont(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    ggml_compute_forward_dup(params, src0, dst);
+    ggml_compute_forward_dup(params, dst);
 }
 
 // ggml_compute_forward_reshape
 
 static void ggml_compute_forward_reshape(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
     // NOP
     UNUSED(params);
-    UNUSED(src0);
     UNUSED(dst);
 }
 
@@ -11099,39 +11265,41 @@ static void ggml_compute_forward_reshape(
 
 static void ggml_compute_forward_view(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0) {
+        const struct ggml_tensor * dst) {
     // NOP
     UNUSED(params);
-    UNUSED(src0);
+    UNUSED(dst);
 }
 
 // ggml_compute_forward_permute
 
 static void ggml_compute_forward_permute(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0) {
+        const struct ggml_tensor * dst) {
     // NOP
     UNUSED(params);
-    UNUSED(src0);
+    UNUSED(dst);
 }
 
 // ggml_compute_forward_transpose
 
 static void ggml_compute_forward_transpose(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0) {
+        const struct ggml_tensor * dst) {
     // NOP
     UNUSED(params);
-    UNUSED(src0);
+    UNUSED(dst);
 }
 
 // ggml_compute_forward_get_rows
 
 static void ggml_compute_forward_get_rows_q(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -11167,9 +11335,11 @@ static void ggml_compute_forward_get_rows_q(
 
 static void ggml_compute_forward_get_rows_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -11202,9 +11372,11 @@ static void ggml_compute_forward_get_rows_f16(
 
 static void ggml_compute_forward_get_rows_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -11237,9 +11409,10 @@ static void ggml_compute_forward_get_rows_f32(
 
 static void ggml_compute_forward_get_rows(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
@@ -11258,16 +11431,16 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
             {
-                ggml_compute_forward_get_rows_q(params, src0, src1, dst);
+                ggml_compute_forward_get_rows_q(params, dst);
             } break;
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_get_rows_f16(params, src0, src1, dst);
+                ggml_compute_forward_get_rows_f16(params, dst);
             } break;
         case GGML_TYPE_F32:
         case GGML_TYPE_I32:
             {
-                ggml_compute_forward_get_rows_f32(params, src0, src1, dst);
+                ggml_compute_forward_get_rows_f32(params, dst);
             } break;
         default:
             {
@@ -11298,9 +11471,11 @@ static void ggml_compute_forward_get_rows(
 
 static void ggml_compute_forward_get_rows_back_f32_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(params->ith == 0);
     GGML_ASSERT(ggml_is_contiguous(dst));
 
@@ -11335,9 +11510,11 @@ static void ggml_compute_forward_get_rows_back_f32_f16(
 
 static void ggml_compute_forward_get_rows_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(params->ith == 0);
     GGML_ASSERT(ggml_is_contiguous(dst));
 
@@ -11372,17 +11549,18 @@ static void ggml_compute_forward_get_rows_back_f32(
 
 static void ggml_compute_forward_get_rows_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_get_rows_back_f32_f16(params, src0, src1, dst);
+                ggml_compute_forward_get_rows_back_f32_f16(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_get_rows_back_f32(params, src0, src1, dst);
+                ggml_compute_forward_get_rows_back_f32(params, dst);
             } break;
         default:
             {
@@ -11413,8 +11591,10 @@ static void ggml_compute_forward_get_rows_back(
 
 static void ggml_compute_forward_diag_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -11453,12 +11633,14 @@ static void ggml_compute_forward_diag_f32(
 
 static void ggml_compute_forward_diag(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_diag_f32(params, src0, dst);
+                ggml_compute_forward_diag_f32(params, dst);
             } break;
         default:
             {
@@ -11471,10 +11653,11 @@ static void ggml_compute_forward_diag(
 
 static void ggml_compute_forward_diag_mask_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst,
         const float value) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -11524,12 +11707,14 @@ static void ggml_compute_forward_diag_mask_f32(
 
 static void ggml_compute_forward_diag_mask_inf(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_diag_mask_f32(params, src0, dst, -INFINITY);
+                ggml_compute_forward_diag_mask_f32(params, dst, -INFINITY);
             } break;
         default:
             {
@@ -11540,12 +11725,14 @@ static void ggml_compute_forward_diag_mask_inf(
 
 static void ggml_compute_forward_diag_mask_zero(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_diag_mask_f32(params, src0, dst, 0);
+                ggml_compute_forward_diag_mask_f32(params, dst, 0);
             } break;
         default:
             {
@@ -11558,10 +11745,12 @@ static void ggml_compute_forward_diag_mask_zero(
 
 static void ggml_compute_forward_soft_max_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * src2,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * src2 = dst->src[2];
+
     assert(ggml_is_contiguous(dst));
     assert(ggml_are_same_shape(src0, dst));
 
@@ -11672,14 +11861,14 @@ static void ggml_compute_forward_soft_max_f32(
 
 static void ggml_compute_forward_soft_max(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * src2,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_soft_max_f32(params, src0, src1, src2, dst);
+                ggml_compute_forward_soft_max_f32(params, dst);
             } break;
         default:
             {
@@ -11692,9 +11881,11 @@ static void ggml_compute_forward_soft_max(
 
 static void ggml_compute_forward_soft_max_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(src1));
     GGML_ASSERT(ggml_is_contiguous(dst));
@@ -11769,13 +11960,14 @@ static void ggml_compute_forward_soft_max_back_f32(
 
 static void ggml_compute_forward_soft_max_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_soft_max_back_f32(params, src0, src1, dst);
+                ggml_compute_forward_soft_max_back_f32(params, dst);
             } break;
         default:
             {
@@ -11788,8 +11980,10 @@ static void ggml_compute_forward_soft_max_back(
 
 static void ggml_compute_forward_alibi_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -11845,8 +12039,10 @@ static void ggml_compute_forward_alibi_f32(
 
 static void ggml_compute_forward_alibi_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -11905,16 +12101,18 @@ static void ggml_compute_forward_alibi_f16(
 
 static void ggml_compute_forward_alibi(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_alibi_f16(params, src0, dst);
+                ggml_compute_forward_alibi_f16(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_alibi_f32(params, src0, dst);
+                ggml_compute_forward_alibi_f32(params, dst);
             } break;
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
@@ -11947,8 +12145,10 @@ static void ggml_compute_forward_alibi(
 
 static void ggml_compute_forward_clamp_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -11987,12 +12187,14 @@ static void ggml_compute_forward_clamp_f32(
 
 static void ggml_compute_forward_clamp(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_clamp_f32(params, src0, dst);
+                ggml_compute_forward_clamp_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_Q4_0:
@@ -12082,10 +12284,12 @@ GGML_CALL void ggml_rope_yarn_corr_dims(
 
 static void ggml_compute_forward_rope_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst,
         const bool forward) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -12258,10 +12462,12 @@ static void ggml_compute_forward_rope_f32(
 
 static void ggml_compute_forward_rope_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst,
         const bool forward) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -12423,17 +12629,18 @@ static void ggml_compute_forward_rope_f16(
 
 static void ggml_compute_forward_rope(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_rope_f16(params, src0, src1, dst, true);
+                ggml_compute_forward_rope_f16(params, dst, true);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rope_f32(params, src0, src1, dst, true);
+                ggml_compute_forward_rope_f32(params, dst, true);
             } break;
         default:
             {
@@ -12446,17 +12653,18 @@ static void ggml_compute_forward_rope(
 
 static void ggml_compute_forward_rope_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_rope_f16(params, src0, src1, dst, false);
+                ggml_compute_forward_rope_f16(params, dst, false);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rope_f32(params, src0, src1, dst, false);
+                ggml_compute_forward_rope_f32(params, dst, false);
             } break;
         default:
             {
@@ -12469,9 +12677,11 @@ static void ggml_compute_forward_rope_back(
 
 static void ggml_compute_forward_conv_transpose_1d_f16_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
@@ -12566,9 +12776,11 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
 
 static void ggml_compute_forward_conv_transpose_1d_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
@@ -12663,17 +12875,18 @@ static void ggml_compute_forward_conv_transpose_1d_f32(
 
 static void ggml_compute_forward_conv_transpose_1d(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_conv_transpose_1d_f16_f32(params, src0, src1, dst);
+                ggml_compute_forward_conv_transpose_1d_f16_f32(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_conv_transpose_1d_f32(params, src0, src1, dst);
+                ggml_compute_forward_conv_transpose_1d_f32(params, dst);
             } break;
         default:
             {
@@ -12687,9 +12900,11 @@ static void ggml_compute_forward_conv_transpose_1d(
 // dst:  result [N, OH, OW, IC*KH*KW]
 static void ggml_compute_forward_im2col_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
@@ -12773,9 +12988,11 @@ static void ggml_compute_forward_im2col_f32(
 // dst:  result [N, OH, OW, IC*KH*KW]
 static void ggml_compute_forward_im2col_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F16);
@@ -12855,17 +13072,15 @@ static void ggml_compute_forward_im2col_f16(
 
 static void ggml_compute_forward_im2col(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
     switch (dst->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_im2col_f16(params, src0, src1, dst);
+                ggml_compute_forward_im2col_f16(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_im2col_f32(params, src0, src1, dst);
+                ggml_compute_forward_im2col_f32(params, dst);
             } break;
         default:
             {
@@ -12879,9 +13094,11 @@ static void ggml_compute_forward_im2col(
 
 static void ggml_compute_forward_conv_transpose_2d(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
@@ -12985,9 +13202,11 @@ static void ggml_compute_forward_conv_transpose_2d(
 static void ggml_compute_forward_pool_1d_sk_p0(
         const struct ggml_compute_params * params,
         const enum ggml_op_pool op,
-        const struct ggml_tensor * src,
         const int k,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src = dst->src[0];
+
     assert(src->type == GGML_TYPE_F32);
     assert(params->ith == 0);
 
@@ -13036,7 +13255,6 @@ static void ggml_compute_forward_pool_1d_sk_p0(
 
 static void ggml_compute_forward_pool_1d(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
               struct ggml_tensor * dst) {
 
     const int32_t * opts = (const int32_t *)dst->op_params;
@@ -13047,15 +13265,17 @@ static void ggml_compute_forward_pool_1d(
     GGML_ASSERT(p0 == 0); // padding not supported
     GGML_ASSERT(k0 == s0); // only s = k supported
 
-    ggml_compute_forward_pool_1d_sk_p0(params, op, src0, k0, dst);
+    ggml_compute_forward_pool_1d_sk_p0(params, op, k0, dst);
 }
 
 // ggml_compute_forward_pool_2d
 
 static void ggml_compute_forward_pool_2d(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src = dst->src[0];
+
     GGML_ASSERT(src->type == GGML_TYPE_F32);
     GGML_ASSERT(params->ith == 0);
 
@@ -13128,9 +13348,10 @@ static void ggml_compute_forward_pool_2d(
 
 static void ggml_compute_forward_upscale_f32(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
     struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -13167,12 +13388,14 @@ static void ggml_compute_forward_upscale_f32(
 
 static void ggml_compute_forward_upscale(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
     struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_upscale_f32(params, src0, dst);
+                ggml_compute_forward_upscale_f32(params, dst);
             } break;
         default:
             {
@@ -13185,9 +13408,10 @@ static void ggml_compute_forward_upscale(
 
 static void ggml_compute_forward_pad_f32(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
           struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -13225,12 +13449,14 @@ static void ggml_compute_forward_pad_f32(
 
 static void ggml_compute_forward_pad(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
     struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_pad_f32(params, src0, dst);
+                ggml_compute_forward_pad_f32(params, dst);
             } break;
         default:
             {
@@ -13243,9 +13469,10 @@ static void ggml_compute_forward_pad(
 
 static void ggml_compute_forward_argsort_f32(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
     struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -13285,13 +13512,14 @@ static void ggml_compute_forward_argsort_f32(
 
 static void ggml_compute_forward_argsort(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
     struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_argsort_f32(params, src0, dst);
+                ggml_compute_forward_argsort_f32(params, dst);
             } break;
         default:
             {
@@ -13304,11 +13532,13 @@ static void ggml_compute_forward_argsort(
 
 static void ggml_compute_forward_flash_attn_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * q,
-        const struct ggml_tensor * k,
-        const struct ggml_tensor * v,
         const bool masked,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * q = dst->src[0];
+    const struct ggml_tensor * k = dst->src[1];
+    const struct ggml_tensor * v = dst->src[2];
+
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
@@ -13494,11 +13724,13 @@ static void ggml_compute_forward_flash_attn_f32(
 
 static void ggml_compute_forward_flash_attn_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * q,
-        const struct ggml_tensor * k,
-        const struct ggml_tensor * v,
         const bool masked,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * q = dst->src[0];
+    const struct ggml_tensor * k = dst->src[1];
+    const struct ggml_tensor * v = dst->src[2];
+
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
@@ -13720,19 +13952,19 @@ static void ggml_compute_forward_flash_attn_f16(
 
 static void ggml_compute_forward_flash_attn(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * q,
-        const struct ggml_tensor * k,
-        const struct ggml_tensor * v,
         const bool masked,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * q = dst->src[0];
+
     switch (q->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_flash_attn_f16(params, q, k, v, masked, dst);
+                ggml_compute_forward_flash_attn_f16(params, masked, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_flash_attn_f32(params, q, k, v, masked, dst);
+                ggml_compute_forward_flash_attn_f32(params, masked, dst);
             } break;
         default:
             {
@@ -13745,12 +13977,14 @@ static void ggml_compute_forward_flash_attn(
 
 static void ggml_compute_forward_flash_ff_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,  // F16
-        const struct ggml_tensor * b0, // F16 fc_w
-        const struct ggml_tensor * b1, // F32 fc_b
-        const struct ggml_tensor * c0, // F16 proj_w
-        const struct ggml_tensor * c1, // F32 proj_b
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * a = dst->src[0];  // F16
+    const struct ggml_tensor * b0 = dst->src[1]; // F16 fc_w
+    const struct ggml_tensor * b1 = dst->src[2]; // F32 fc_b
+    const struct ggml_tensor * c0 = dst->src[3]; // F16 proj_w
+    const struct ggml_tensor * c1 = dst->src[4]; // F32 proj_b
+
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
@@ -13878,16 +14112,14 @@ static void ggml_compute_forward_flash_ff_f16(
 
 static void ggml_compute_forward_flash_ff(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b0,
-        const struct ggml_tensor * b1,
-        const struct ggml_tensor * c0,
-        const struct ggml_tensor * c1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * b0 = dst->src[1];
+
     switch (b0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_flash_ff_f16(params, a, b0, b1, c0, c1, dst);
+                ggml_compute_forward_flash_ff_f16(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
@@ -13904,12 +14136,14 @@ static void ggml_compute_forward_flash_ff(
 
 static void ggml_compute_forward_flash_attn_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * q,
-        const struct ggml_tensor * k,
-        const struct ggml_tensor * v,
-        const struct ggml_tensor * d,
         const bool masked,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * q = dst->src[0];
+    const struct ggml_tensor * k = dst->src[1];
+    const struct ggml_tensor * v = dst->src[2];
+    const struct ggml_tensor * d = dst->src[3];
+
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
@@ -14257,16 +14491,15 @@ static void ggml_compute_forward_flash_attn_back_f32(
 
 static void ggml_compute_forward_flash_attn_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * q,
-        const struct ggml_tensor * k,
-        const struct ggml_tensor * v,
-        const struct ggml_tensor * d,
         const bool masked,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * q = dst->src[0];
+
     switch (q->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_flash_attn_back_f32(params, q, k, v, d, masked, dst);
+                ggml_compute_forward_flash_attn_back_f32(params, masked, dst);
             } break;
         default:
             {
@@ -14279,8 +14512,10 @@ static void ggml_compute_forward_flash_attn_back(
 
 static void ggml_compute_forward_win_part_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -14323,12 +14558,14 @@ static void ggml_compute_forward_win_part_f32(
 
 static void ggml_compute_forward_win_part(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_win_part_f32(params, src0, dst);
+                ggml_compute_forward_win_part_f32(params, dst);
             } break;
         default:
             {
@@ -14341,8 +14578,10 @@ static void ggml_compute_forward_win_part(
 
 static void ggml_compute_forward_win_unpart_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -14383,12 +14622,14 @@ static void ggml_compute_forward_win_unpart_f32(
 
 static void ggml_compute_forward_win_unpart(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_win_unpart_f32(params, src0, dst);
+                ggml_compute_forward_win_unpart_f32(params, dst);
             } break;
         default:
             {
@@ -14401,58 +14642,58 @@ static void ggml_compute_forward_win_unpart(
 
 static void ggml_compute_forward_unary(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
     const enum ggml_unary_op op = ggml_get_unary_op(dst);
 
     switch (op) {
         case GGML_UNARY_OP_ABS:
             {
-                ggml_compute_forward_abs(params, src0, dst);
+                ggml_compute_forward_abs(params, dst);
             } break;
         case GGML_UNARY_OP_SGN:
             {
-                ggml_compute_forward_sgn(params, src0, dst);
+                ggml_compute_forward_sgn(params, dst);
             } break;
         case GGML_UNARY_OP_NEG:
             {
-                ggml_compute_forward_neg(params, src0, dst);
+                ggml_compute_forward_neg(params, dst);
             } break;
         case GGML_UNARY_OP_STEP:
             {
-                ggml_compute_forward_step(params, src0, dst);
+                ggml_compute_forward_step(params, dst);
             } break;
         case GGML_UNARY_OP_TANH:
             {
-                ggml_compute_forward_tanh(params, src0, dst);
+                ggml_compute_forward_tanh(params, dst);
             } break;
         case GGML_UNARY_OP_ELU:
             {
-                ggml_compute_forward_elu(params, src0, dst);
+                ggml_compute_forward_elu(params, dst);
             } break;
         case GGML_UNARY_OP_RELU:
             {
-                ggml_compute_forward_relu(params, src0, dst);
+                ggml_compute_forward_relu(params, dst);
             } break;
         case GGML_UNARY_OP_GELU:
             {
-                ggml_compute_forward_gelu(params, src0, dst);
+                ggml_compute_forward_gelu(params, dst);
             } break;
         case GGML_UNARY_OP_GELU_QUICK:
             {
-                ggml_compute_forward_gelu_quick(params, src0, dst);
+                ggml_compute_forward_gelu_quick(params, dst);
             } break;
         case GGML_UNARY_OP_SILU:
             {
-                ggml_compute_forward_silu(params, src0, dst);
+                ggml_compute_forward_silu(params, dst);
             } break;
         case GGML_UNARY_OP_HARDSWISH:
             {
-                ggml_compute_forward_hardswish(params, src0, dst);
+                ggml_compute_forward_hardswish(params, dst);
             } break;
         case GGML_UNARY_OP_HARDSIGMOID:
             {
-                ggml_compute_forward_hardsigmoid(params, src0, dst);
+                ggml_compute_forward_hardsigmoid(params, dst);
             } break;
         default:
             {
@@ -14465,8 +14706,10 @@ static void ggml_compute_forward_unary(
 
 static void ggml_compute_forward_get_rel_pos_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -14492,12 +14735,14 @@ static void ggml_compute_forward_get_rel_pos_f16(
 
 static void ggml_compute_forward_get_rel_pos(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_get_rel_pos_f16(params, src0, dst);
+                ggml_compute_forward_get_rel_pos_f16(params, dst);
             } break;
         default:
             {
@@ -14510,11 +14755,12 @@ static void ggml_compute_forward_get_rel_pos(
 
 static void ggml_compute_forward_add_rel_pos_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * src2,
         struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * src2 = dst->src[2];
+
     const bool inplace = (bool) ((int32_t *) dst->op_params)[0];
     if (!inplace && params->type == GGML_TASK_INIT) {
         if (params->ith != 0) {
@@ -14578,14 +14824,14 @@ static void ggml_compute_forward_add_rel_pos_f32(
 
 static void ggml_compute_forward_add_rel_pos(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * src2,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_add_rel_pos_f32(params, src0, src1, src2, dst);
+                ggml_compute_forward_add_rel_pos_f32(params, dst);
             } break;
         default:
             {
@@ -14598,9 +14844,11 @@ static void ggml_compute_forward_add_rel_pos(
 
 static void ggml_compute_forward_map_unary_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst,
         const ggml_unary_op_f32_t fun) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -14622,13 +14870,15 @@ static void ggml_compute_forward_map_unary_f32(
 
 static void ggml_compute_forward_map_unary(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst,
         const ggml_unary_op_f32_t fun) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_map_unary_f32(params, src0, dst, fun);
+                ggml_compute_forward_map_unary_f32(params, dst, fun);
             } break;
         default:
             {
@@ -14641,10 +14891,12 @@ static void ggml_compute_forward_map_unary(
 
 static void ggml_compute_forward_map_binary_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst,
         const ggml_binary_op_f32_t fun) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     assert(params->ith == 0);
     assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
@@ -14669,14 +14921,15 @@ static void ggml_compute_forward_map_binary_f32(
 
 static void ggml_compute_forward_map_binary(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst,
         const ggml_binary_op_f32_t fun) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_map_binary_f32(params, src0, src1, dst, fun);
+                ggml_compute_forward_map_binary_f32(params, dst, fun);
             } break;
         default:
             {
@@ -14689,9 +14942,11 @@ static void ggml_compute_forward_map_binary(
 
 static void ggml_compute_forward_map_custom1_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
         struct ggml_tensor * dst,
         const ggml_custom1_op_f32_t fun) {
+
+    const struct ggml_tensor * a = dst->src[0];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -14705,10 +14960,12 @@ static void ggml_compute_forward_map_custom1_f32(
 
 static void ggml_compute_forward_map_custom2_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b,
         struct ggml_tensor * dst,
         const ggml_custom2_op_f32_t fun) {
+
+    const struct ggml_tensor * a = dst->src[0];
+    const struct ggml_tensor * b = dst->src[1];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -14722,11 +14979,13 @@ static void ggml_compute_forward_map_custom2_f32(
 
 static void ggml_compute_forward_map_custom3_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b,
-        const struct ggml_tensor * c,
         struct ggml_tensor * dst,
         const ggml_custom3_op_f32_t fun) {
+
+    const struct ggml_tensor * a = dst->src[0];
+    const struct ggml_tensor * b = dst->src[1];
+    const struct ggml_tensor * c = dst->src[1];
+
     assert(params->ith == 0);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -14740,8 +14999,10 @@ static void ggml_compute_forward_map_custom3_f32(
 
 static void ggml_compute_forward_map_custom1(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * a = dst->src[0];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -14755,9 +15016,11 @@ static void ggml_compute_forward_map_custom1(
 
 static void ggml_compute_forward_map_custom2(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * a = dst->src[0];
+    const struct ggml_tensor * b = dst->src[1];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -14771,10 +15034,12 @@ static void ggml_compute_forward_map_custom2(
 
 static void ggml_compute_forward_map_custom3(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b,
-        const struct ggml_tensor * c,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * a = dst->src[0];
+    const struct ggml_tensor * b = dst->src[1];
+    const struct ggml_tensor * c = dst->src[2];
+
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -14788,9 +15053,11 @@ static void ggml_compute_forward_map_custom3(
 
 static void ggml_compute_forward_cross_entropy_loss_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(src1));
     GGML_ASSERT(ggml_is_scalar(dst));
@@ -14894,13 +15161,14 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
 
 static void ggml_compute_forward_cross_entropy_loss(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_cross_entropy_loss_f32(params, src0, src1, dst);
+                ggml_compute_forward_cross_entropy_loss_f32(params, dst);
             } break;
         default:
             {
@@ -14913,10 +15181,12 @@ static void ggml_compute_forward_cross_entropy_loss(
 
 static void ggml_compute_forward_cross_entropy_loss_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * opt0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * opt0 = dst->src[2];
+
     GGML_ASSERT(ggml_is_contiguous(dst));
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(src1));
@@ -15003,14 +15273,14 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
 
 static void ggml_compute_forward_cross_entropy_loss_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * opt0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_cross_entropy_loss_back_f32(params, src0, src1, opt0, dst);
+                ggml_compute_forward_cross_entropy_loss_back_f32(params, dst);
             } break;
         default:
             {
@@ -15058,312 +15328,312 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
     switch (tensor->op) {
         case GGML_OP_DUP:
             {
-                ggml_compute_forward_dup(params, tensor->src[0], tensor);
+                ggml_compute_forward_dup(params, tensor);
             } break;
         case GGML_OP_ADD:
             {
-                ggml_compute_forward_add(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_add(params, tensor);
             } break;
         case GGML_OP_ADD1:
             {
-                ggml_compute_forward_add1(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_add1(params, tensor);
             } break;
         case GGML_OP_ACC:
             {
-                ggml_compute_forward_acc(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_acc(params, tensor);
             } break;
         case GGML_OP_SUB:
             {
-                ggml_compute_forward_sub(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_sub(params, tensor);
             } break;
         case GGML_OP_MUL:
             {
-                ggml_compute_forward_mul(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_mul(params, tensor);
             } break;
         case GGML_OP_DIV:
             {
-                ggml_compute_forward_div(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_div(params, tensor);
             } break;
         case GGML_OP_SQR:
             {
-                ggml_compute_forward_sqr(params, tensor->src[0], tensor);
+                ggml_compute_forward_sqr(params, tensor);
             } break;
         case GGML_OP_SQRT:
             {
-                ggml_compute_forward_sqrt(params, tensor->src[0], tensor);
+                ggml_compute_forward_sqrt(params, tensor);
             } break;
         case GGML_OP_LOG:
             {
-                ggml_compute_forward_log(params, tensor->src[0], tensor);
+                ggml_compute_forward_log(params, tensor);
             } break;
         case GGML_OP_SUM:
             {
-                ggml_compute_forward_sum(params, tensor->src[0], tensor);
+                ggml_compute_forward_sum(params, tensor);
             } break;
         case GGML_OP_SUM_ROWS:
             {
-                ggml_compute_forward_sum_rows(params, tensor->src[0], tensor);
+                ggml_compute_forward_sum_rows(params, tensor);
             } break;
         case GGML_OP_MEAN:
             {
-                ggml_compute_forward_mean(params, tensor->src[0], tensor);
+                ggml_compute_forward_mean(params, tensor);
             } break;
         case GGML_OP_ARGMAX:
             {
-                ggml_compute_forward_argmax(params, tensor->src[0], tensor);
+                ggml_compute_forward_argmax(params, tensor);
             } break;
         case GGML_OP_REPEAT:
             {
-                ggml_compute_forward_repeat(params, tensor->src[0], tensor);
+                ggml_compute_forward_repeat(params, tensor);
             } break;
         case GGML_OP_REPEAT_BACK:
             {
-                ggml_compute_forward_repeat_back(params, tensor->src[0], tensor);
+                ggml_compute_forward_repeat_back(params, tensor);
             } break;
         case GGML_OP_CONCAT:
             {
-                ggml_compute_forward_concat(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_concat(params, tensor);
             } break;
         case GGML_OP_SILU_BACK:
             {
-                ggml_compute_forward_silu_back(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_silu_back(params, tensor);
             } break;
         case GGML_OP_NORM:
             {
-                ggml_compute_forward_norm(params, tensor->src[0], tensor);
+                ggml_compute_forward_norm(params, tensor);
             } break;
         case GGML_OP_RMS_NORM:
             {
-                ggml_compute_forward_rms_norm(params, tensor->src[0], tensor);
+                ggml_compute_forward_rms_norm(params, tensor);
             } break;
         case GGML_OP_RMS_NORM_BACK:
             {
-                ggml_compute_forward_rms_norm_back(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_rms_norm_back(params, tensor);
             } break;
         case GGML_OP_GROUP_NORM:
             {
-                ggml_compute_forward_group_norm(params, tensor->src[0], tensor);
+                ggml_compute_forward_group_norm(params, tensor);
             } break;
         case GGML_OP_MUL_MAT:
             {
-                ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_mul_mat(params, tensor);
             } break;
         case GGML_OP_MUL_MAT_ID:
             {
-                ggml_compute_forward_mul_mat_id(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_mul_mat_id(params, tensor);
             } break;
         case GGML_OP_OUT_PROD:
             {
-                ggml_compute_forward_out_prod(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_out_prod(params, tensor);
             } break;
         case GGML_OP_SCALE:
             {
-                ggml_compute_forward_scale(params, tensor->src[0], tensor);
+                ggml_compute_forward_scale(params, tensor);
             } break;
         case GGML_OP_SET:
             {
-                ggml_compute_forward_set(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_set(params, tensor);
             } break;
         case GGML_OP_CPY:
             {
-                ggml_compute_forward_cpy(params, tensor->src[0], tensor);
+                ggml_compute_forward_cpy(params, tensor);
             } break;
         case GGML_OP_CONT:
             {
-                ggml_compute_forward_cont(params, tensor->src[0], tensor);
+                ggml_compute_forward_cont(params, tensor);
             } break;
         case GGML_OP_RESHAPE:
             {
-                ggml_compute_forward_reshape(params, tensor->src[0], tensor);
+                ggml_compute_forward_reshape(params, tensor);
             } break;
         case GGML_OP_VIEW:
             {
-                ggml_compute_forward_view(params, tensor->src[0]);
+                ggml_compute_forward_view(params, tensor);
             } break;
         case GGML_OP_PERMUTE:
             {
-                ggml_compute_forward_permute(params, tensor->src[0]);
+                ggml_compute_forward_permute(params, tensor);
             } break;
         case GGML_OP_TRANSPOSE:
             {
-                ggml_compute_forward_transpose(params, tensor->src[0]);
+                ggml_compute_forward_transpose(params, tensor);
             } break;
         case GGML_OP_GET_ROWS:
             {
-                ggml_compute_forward_get_rows(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_get_rows(params, tensor);
             } break;
         case GGML_OP_GET_ROWS_BACK:
             {
-                ggml_compute_forward_get_rows_back(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_get_rows_back(params, tensor);
             } break;
         case GGML_OP_DIAG:
             {
-                ggml_compute_forward_diag(params, tensor->src[0], tensor);
+                ggml_compute_forward_diag(params, tensor);
             } break;
         case GGML_OP_DIAG_MASK_INF:
             {
-                ggml_compute_forward_diag_mask_inf(params, tensor->src[0], tensor);
+                ggml_compute_forward_diag_mask_inf(params, tensor);
             } break;
         case GGML_OP_DIAG_MASK_ZERO:
             {
-                ggml_compute_forward_diag_mask_zero(params, tensor->src[0], tensor);
+                ggml_compute_forward_diag_mask_zero(params, tensor);
             } break;
         case GGML_OP_SOFT_MAX:
             {
-                ggml_compute_forward_soft_max(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
+                ggml_compute_forward_soft_max(params, tensor);
             } break;
         case GGML_OP_SOFT_MAX_BACK:
             {
-                ggml_compute_forward_soft_max_back(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_soft_max_back(params, tensor);
             } break;
         case GGML_OP_ROPE:
             {
-                ggml_compute_forward_rope(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_rope(params, tensor);
             } break;
         case GGML_OP_ROPE_BACK:
             {
-                ggml_compute_forward_rope_back(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_rope_back(params, tensor);
             } break;
         case GGML_OP_ALIBI:
             {
-                ggml_compute_forward_alibi(params, tensor->src[0], tensor);
+                ggml_compute_forward_alibi(params, tensor);
             } break;
         case GGML_OP_CLAMP:
             {
-                ggml_compute_forward_clamp(params, tensor->src[0], tensor);
+                ggml_compute_forward_clamp(params, tensor);
             } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
-                ggml_compute_forward_conv_transpose_1d(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_conv_transpose_1d(params, tensor);
             } break;
         case GGML_OP_IM2COL:
             {
-                ggml_compute_forward_im2col(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_im2col(params, tensor);
             } break;
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
-                ggml_compute_forward_conv_transpose_2d(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_conv_transpose_2d(params, tensor);
             } break;
         case GGML_OP_POOL_1D:
             {
-                ggml_compute_forward_pool_1d(params, tensor->src[0], tensor);
+                ggml_compute_forward_pool_1d(params, tensor);
             } break;
         case GGML_OP_POOL_2D:
             {
-                ggml_compute_forward_pool_2d(params, tensor->src[0], tensor);
+                ggml_compute_forward_pool_2d(params, tensor);
             } break;
         case GGML_OP_UPSCALE:
             {
-                ggml_compute_forward_upscale(params, tensor->src[0], tensor);
+                ggml_compute_forward_upscale(params, tensor);
             } break;
         case GGML_OP_PAD:
             {
-                ggml_compute_forward_pad(params, tensor->src[0], tensor);
+                ggml_compute_forward_pad(params, tensor);
             } break;
         case GGML_OP_ARGSORT:
             {
-                ggml_compute_forward_argsort(params, tensor->src[0], tensor);
+                ggml_compute_forward_argsort(params, tensor);
             } break;
         case GGML_OP_LEAKY_RELU:
             {
-                ggml_compute_forward_leaky_relu(params, tensor->src[0], tensor);
+                ggml_compute_forward_leaky_relu(params, tensor);
             } break;
         case GGML_OP_FLASH_ATTN:
             {
                 const int32_t t = ggml_get_op_params_i32(tensor, 0);
                 GGML_ASSERT(t == 0 || t == 1);
                 const bool masked = t != 0;
-                ggml_compute_forward_flash_attn(params, tensor->src[0], tensor->src[1], tensor->src[2], masked, tensor);
+                ggml_compute_forward_flash_attn(params, masked, tensor);
             } break;
         case GGML_OP_FLASH_FF:
             {
-                ggml_compute_forward_flash_ff(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], tensor->src[4], tensor);
+                ggml_compute_forward_flash_ff(params, tensor);
             } break;
         case GGML_OP_FLASH_ATTN_BACK:
             {
                 int32_t t = ggml_get_op_params_i32(tensor, 0);
                 GGML_ASSERT(t == 0 || t == 1);
                 bool masked = t != 0;
-                ggml_compute_forward_flash_attn_back(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], masked, tensor);
+                ggml_compute_forward_flash_attn_back(params, masked, tensor);
             } break;
         case GGML_OP_WIN_PART:
             {
-                ggml_compute_forward_win_part(params, tensor->src[0], tensor);
+                ggml_compute_forward_win_part(params, tensor);
             } break;
         case GGML_OP_WIN_UNPART:
             {
-                ggml_compute_forward_win_unpart(params, tensor->src[0], tensor);
+                ggml_compute_forward_win_unpart(params, tensor);
             } break;
         case GGML_OP_UNARY:
             {
-                ggml_compute_forward_unary(params, tensor->src[0], tensor);
+                ggml_compute_forward_unary(params, tensor);
             } break;
         case GGML_OP_GET_REL_POS:
             {
-                ggml_compute_forward_get_rel_pos(params, tensor->src[0], tensor);
+                ggml_compute_forward_get_rel_pos(params, tensor);
             } break;
         case GGML_OP_ADD_REL_POS:
             {
-                ggml_compute_forward_add_rel_pos(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
+                ggml_compute_forward_add_rel_pos(params, tensor);
             } break;
         case GGML_OP_MAP_UNARY:
             {
                 ggml_unary_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_unary(params, tensor->src[0], tensor, fun);
+                ggml_compute_forward_map_unary(params, tensor, fun);
             }
             break;
         case GGML_OP_MAP_BINARY:
             {
                 ggml_binary_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_binary(params, tensor->src[0], tensor->src[1], tensor, fun);
+                ggml_compute_forward_map_binary(params, tensor, fun);
             }
             break;
         case GGML_OP_MAP_CUSTOM1_F32:
             {
                 ggml_custom1_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom1_f32(params, tensor->src[0], tensor, fun);
+                ggml_compute_forward_map_custom1_f32(params, tensor, fun);
             }
             break;
         case GGML_OP_MAP_CUSTOM2_F32:
             {
                 ggml_custom2_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom2_f32(params, tensor->src[0], tensor->src[1], tensor, fun);
+                ggml_compute_forward_map_custom2_f32(params, tensor, fun);
             }
             break;
         case GGML_OP_MAP_CUSTOM3_F32:
             {
                 ggml_custom3_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom3_f32(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor, fun);
+                ggml_compute_forward_map_custom3_f32(params, tensor, fun);
             }
             break;
         case GGML_OP_MAP_CUSTOM1:
             {
-                ggml_compute_forward_map_custom1(params, tensor->src[0], tensor);
+                ggml_compute_forward_map_custom1(params, tensor);
             }
             break;
         case GGML_OP_MAP_CUSTOM2:
             {
-                ggml_compute_forward_map_custom2(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_map_custom2(params, tensor);
             }
             break;
         case GGML_OP_MAP_CUSTOM3:
             {
-                ggml_compute_forward_map_custom3(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
+                ggml_compute_forward_map_custom3(params, tensor);
             }
             break;
         case GGML_OP_CROSS_ENTROPY_LOSS:
             {
-                ggml_compute_forward_cross_entropy_loss(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_cross_entropy_loss(params, tensor);
             }
             break;
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
             {
-                ggml_compute_forward_cross_entropy_loss_back(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
+                ggml_compute_forward_cross_entropy_loss_back(params, tensor);
             }
             break;
         case GGML_OP_NONE:
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 114a9a974..8f9139d1b 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -111,6 +111,7 @@ class MODEL_ARCH(IntEnum):
     ORION      = auto()
     INTERNLM2  = auto()
     MINICPM    = auto()
+    GEMMA      = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -167,6 +168,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.ORION:          "orion",
     MODEL_ARCH.INTERNLM2:      "internlm2",
     MODEL_ARCH.MINICPM:        "minicpm",
+    MODEL_ARCH.GEMMA:          "gemma",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -511,6 +513,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
+    MODEL_ARCH.GEMMA: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_NORM,
+    ],
     # TODO
 }
 
diff --git a/llama.cpp b/llama.cpp
index faff13e21..859d909c0 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -232,6 +232,7 @@ enum llm_arch {
     LLM_ARCH_ORION,
     LLM_ARCH_INTERNLM2,
     LLM_ARCH_MINICPM,
+    LLM_ARCH_GEMMA,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -258,6 +259,7 @@ static std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_ORION,           "orion"      },
     { LLM_ARCH_INTERNLM2,       "internlm2"  },
     { LLM_ARCH_MINICPM,         "minicpm"    },
+    { LLM_ARCH_GEMMA,           "gemma"      },
 };
 
 enum llm_kv {
@@ -784,6 +786,22 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
             { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
         },
     },
+    {
+        LLM_ARCH_GEMMA,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2806,13 +2824,7 @@ struct llama_model_loader {
 
         std::vector<no_init<uint8_t>> read_buf;
 
-        for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) {
-            struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i));
-            if (!cur) {
-                // some tensors may be allocated in a different context
-                continue;
-            }
-
+        for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) {
             if (progress_callback) {
                 if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) {
                     return false;
@@ -3289,6 +3301,16 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GEMMA:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 18: model.type = e_model::MODEL_2B; break;
+                    case 28: model.type = e_model::MODEL_7B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+               }
+            } break;
         default: (void)0;
     }
 
@@ -3766,7 +3788,7 @@ static bool llm_load_tensors(
     }
 
     // create one context per buffer type
-    size_t ctx_size = ggml_tensor_overhead()*ml.n_tensors;
+    size_t ctx_size = ggml_tensor_overhead()*(ml.n_tensors + 1); // +1 for models where tok_embd is duplicated as output
     std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     for (auto & it : buft_layer_count) {
         struct ggml_init_params params = {
@@ -3904,6 +3926,7 @@ static bool llm_load_tensors(
                         } else {
                             model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
                             ml.n_created--; // artificial tensor
+                            ml.size_data += ggml_nbytes(model.output);
                         }
                     }
 
@@ -4432,6 +4455,40 @@ static bool llm_load_tensors(
                         layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
                     }
                 } break;
+            case LLM_ARCH_GEMMA:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading
+                    ml.n_created--; // artificial tensor
+                    ml.size_data += ggml_nbytes(model.output);
+
+                    const int64_t n_ff          = hparams.n_ff;
+                    const int64_t n_embd_head_k = hparams.n_embd_head_k;
+                    const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
+                    const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
+
+                    for (uint32_t i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * hparams.n_head});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -7438,6 +7495,113 @@ struct llm_build_context {
 
         return gf;
     }
+
+    struct ggml_cgraph * build_gemma() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
+        cb(inpL, "inp_embd", -1);
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
+        cb(inp_pos, "inp_pos", -1);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
+        cb(KQ_mask, "KQ_mask", -1);
+
+        // shift the entire K-cache if needed
+        if (do_rope_shift) {
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
+        }
+
+        for (int il = 0; il < n_layer; ++il) {
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_custom(
+                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos,
+                        n_embd_head_k, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur", il);
+                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
+                cb(Qcur, "Qcur_scaled", il);
+
+                Kcur = ggml_rope_custom(
+                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos,
+                        n_embd_head_k, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+                cb(cur, "kqv_out", il);
+            }
+            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+            cb(sa_out, "sa_out", il);
+
+            cur = llm_build_norm(ctx0, sa_out, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_ffn(ctx0, cur,
+                        model.layers[il].ffn_up, NULL,
+                        model.layers[il].ffn_gate, NULL,
+                        model.layers[il].ffn_down, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, sa_out);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
 };
 
 static struct ggml_cgraph * llama_build_graph(
@@ -7546,6 +7710,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_minicpm();
             } break;
+        case LLM_ARCH_GEMMA:
+            {
+                result = llm.build_gemma();
+            } break;
         default:
             GGML_ASSERT(false);
     }
@@ -12317,18 +12485,19 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
         data_ctx->write(&kv_used,     sizeof(kv_used));
 
         if (kv_buf_size) {
-            const size_t elt_size = ggml_element_size(kv_self.k_l[0]);
-
             std::vector<uint8_t> tmp_buf;
             for (int il = 0; il < (int) n_layer; ++il) {
-                tmp_buf.resize(elt_size*n_embd_k_gqa*kv_head);
+                size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head);
+                tmp_buf.resize(k_size);
                 ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), 0, tmp_buf.size());
                 data_ctx->write(tmp_buf.data(), tmp_buf.size());
 
                 // v is not contiguous, copy row by row
-                tmp_buf.resize(elt_size*kv_head);
+                size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head);
+                size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, n_ctx);
+                tmp_buf.resize(v_row_size);
                 for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) {
-                    ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), ir*elt_size*n_ctx, tmp_buf.size());
+                    ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), ir*v_row_stride, tmp_buf.size());
                     data_ctx->write(tmp_buf.data(), tmp_buf.size());
                 }
             }
@@ -12430,17 +12599,16 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
         if (kv_buf_size) {
             GGML_ASSERT(kv_self.total_size() == kv_buf_size);
 
-            const size_t elt_size = ggml_element_size(kv_self.k_l[0]);
-
             for (int il = 0; il < (int) n_layer; ++il) {
-                size_t k_size = elt_size*n_embd_k_gqa*kv_head;
+                size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head);
                 ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size);
                 inp += k_size;
 
                 // v is not contiguous, copy row by row
-                size_t v_row_size = elt_size*kv_head;
+                size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head);
+                size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, n_ctx);
                 for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) {
-                    ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*elt_size*n_ctx, v_row_size);
+                    ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*v_row_stride, v_row_size);
                     inp += v_row_size;
                 }
             }
diff --git a/llama.h b/llama.h
index b4f7a4f07..ef8667456 100644
--- a/llama.h
+++ b/llama.h
@@ -710,7 +710,7 @@ extern "C" {
 
     /// Apply chat template. Inspired by hf apply_chat_template() on python.
     /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
-    /// NOTE: This function only support some known jinja templates. It is not a jinja parser.
+    /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
     /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
     /// @param chat Pointer to a list of multiple llama_chat_message
     /// @param n_msg Number of llama_chat_message in this chat