model : Granite MoE shared (#13269)

* feat: Add GGUF conversion for granitemoeshared

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: hparam and arch plumbing for granitemoeshared

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Split MoE fused tensors for shared experts in conversion

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: First WIP cut at model arch in cpp

The hparam and architecture plumbing should be correct, but the
implementation of the shared experts seems to still be broken.

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Cleaner (maybe more correct?) splitting for gate/up

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Fix the input to the shared experts

I had misread that the shared experts take the inputs _before_ the standard
MoE layer and was feeding the output of the MoE to the shared experts.

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Avoid architecture-specific checks for Granite MoE Shared

This is a cleaner way that will allow more flexibility in architecture
strings going forward.

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* refactor: Split granite architectures out of llm_build_llama

This helps de-clutter the llama-family graph construction and allows
granite to diverge further (in preparation for Granite 4).

NOTE: I removed the granite scale factors from llm_build_deci because they
appear to only be there as copy-paste from llm_build_llama. The HF config
does not seem to set those values:
https://huggingface.co/Deci/DeciLM-7B/blob/main/config.json

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Fix compiler warning about uninitialized inp_pos

This should not have been reachable, but it warns on some compliers

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Consoladate GraniteMoEShared into GraniteMoE for conversion

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Consolidate GraniteMoEShared into GraniteMoE on the c++ side

Branch: GraniteMoEShared

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

---------

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

build-xcframework.sh

@@ -117,6 +117,7 @@ setup_framework_structure() {
     # Copy all required headers (common for all platforms)
     cp include/llama.h             ${header_path}
     cp ggml/include/ggml.h         ${header_path}
+    cp ggml/include/ggml-opt.h     ${header_path}
     cp ggml/include/ggml-alloc.h   ${header_path}
     cp ggml/include/ggml-backend.h ${header_path}
     cp ggml/include/ggml-metal.h   ${header_path}

common/common.cpp

@@ -1565,3 +1565,20 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 
     return result;
 }
+
+ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride) {
+    const int64_t ne_datapoint = llama_n_ctx(ctx);
+    const int64_t ndata        = (tokens.size() - ne_datapoint - 1) / stride;
+    ggml_opt_dataset_t result = ggml_opt_dataset_init(
+        GGML_TYPE_I32, GGML_TYPE_I32, ne_datapoint, ne_datapoint, ndata, /*ndata_shard =*/ 1);
+
+    llama_token * data   = (llama_token *) ggml_opt_dataset_data(result)->data;
+    llama_token * labels = (llama_token *) ggml_opt_dataset_labels(result)->data;
+
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        memcpy(data   + idata*ne_datapoint, tokens.data() + idata*stride + 0, ne_datapoint*sizeof(llama_token));
+        memcpy(labels + idata*ne_datapoint, tokens.data() + idata*stride + 1, ne_datapoint*sizeof(llama_token));
+    }
+
+    return result;
+}

common/common.h

@@ -666,3 +666,9 @@ const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 
 }
+
+//
+// training utils
+//
+
+ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);

convert_hf_to_gguf.py

@@ -5746,11 +5746,20 @@ class GraniteModel(LlamaModel):
             logger.info("gguf: (granite) logits_scale = %s", logits_scale)
 
 
-@ModelBase.register("GraniteMoeForCausalLM")
+@ModelBase.register("GraniteMoeForCausalLM", "GraniteMoeSharedForCausalLM")
 class GraniteMoeModel(GraniteModel):
     """Conversion for IBM's GraniteMoeForCausalLM"""
     model_arch = gguf.MODEL_ARCH.GRANITE_MOE
 
+    def set_gguf_parameters(self):
+        """GraniteMoeShared uses GraniteMoe parameters plus the following:
+        - shared_intermediate_size
+        """
+        super().set_gguf_parameters()
+        if shared_feed_forward_length := self.hparams.get("shared_intermediate_size"):
+            self.gguf_writer.add_expert_shared_feed_forward_length(shared_feed_forward_length)
+            logger.info("gguf: (granitemoeshared) shared_feed_forward_length = %s", shared_feed_forward_length)
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         """In modeling_granitemoe, the JetMoe implementation of parallel experts
         is used. This essentially merges w1 and w3 into a single tensor with 2x
@@ -5761,12 +5770,21 @@ class GraniteMoeModel(GraniteModel):
         if name.endswith("block_sparse_moe.input_linear.weight"):
             ffn_dim = self.hparams["intermediate_size"]
             assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
-            gate, up = data_torch[..., :ffn_dim, :], data_torch[..., ffn_dim:, :]
+            gate, up = data_torch.split(ffn_dim, dim=-2)
             return [
                 (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
                 (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
             ]
 
+        if name.endswith("shared_mlp.input_linear.weight"):
+            ffn_dim = self.hparams["shared_intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
+            gate, up = data_torch.split(ffn_dim, dim=-2)
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+            ]
+
         return super().modify_tensors(data_torch, name, bid)
 
 

examples/CMakeLists.txt

@@ -32,6 +32,7 @@ else()
     add_subdirectory(speculative)
     add_subdirectory(speculative-simple)
     add_subdirectory(gen-docs)
+    add_subdirectory(training)
     if (NOT GGML_BACKEND_DL)
         add_subdirectory(convert-llama2c-to-ggml)
         # these examples use the backends directly and cannot be built with dynamic loading

examples/training/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-finetune)
+add_executable(${TARGET} finetune.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/training/README.md

@@ -0,0 +1,17 @@
+# llama.cpp/examples/training
+
+This directory contains examples related to language model training using llama.cpp/GGML.
+So far finetuning is technically functional (for FP32 models and limited hardware setups) but the code is very much WIP.
+Finetuning of Stories 260K and LLaMA 3.2 1b seems to work with 24 GB of memory.
+**For CPU training, compile llama.cpp without any additional backends such as CUDA.**
+**For CUDA training, use the maximum number of GPU layers.**
+
+Proof of concept:
+
+``` sh
+export model_name=llama_3.2-1b && export quantization=f32
+./build/bin/finetune --file wikitext-2-raw/wiki.test.raw -ngl 999 --model models/${model_name}-${quantization}.gguf -c 512 -b 512 -ub 512
+./build/bin/perplexity --file wikitext-2-raw/wiki.test.raw -ngl 999 --model finetuned-model.gguf
+```
+
+The perplexity value of the finetuned model should be lower after training on the test set for 2 epochs.

examples/training/finetune.cpp

@@ -0,0 +1,96 @@
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    params.escape = false;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
+        return 1;
+    }
+
+    if (params.use_mmap) {
+        LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n", __func__);
+        params.use_mmap = false;
+    }
+    if (params.cache_type_k != GGML_TYPE_F32) {
+        LOG_INF("%s: force changing k cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__);
+        params.cache_type_k = GGML_TYPE_F32;
+    }
+    if (params.cache_type_v != GGML_TYPE_F32) {
+        LOG_INF("%s: force changing v cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__);
+        params.cache_type_v = GGML_TYPE_F32;
+    }
+
+    common_init();
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // load the model and apply lora adapter, if any
+    common_init_result llama_init = common_init_from_params(params);
+    llama_model_ptr   & model = llama_init.model;
+    llama_context_ptr & ctx   = llama_init.context;
+
+    if (model == NULL) {
+        LOG_ERR("%s: unable to load model\n", __func__);
+        return 1;
+    }
+
+    // print system information
+    {
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+    }
+
+    constexpr float val_split = 0.05f;
+
+    std::vector<llama_token> tokens = common_tokenize(ctx.get(), params.prompt, true);
+    ggml_opt_dataset_t dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get())/2);
+
+    struct ggml_opt_optimizer_params optimizer_params = ggml_opt_get_default_optimizer_params(nullptr);
+    optimizer_params.adamw.alpha = 1e-7f; // learning rate
+
+    struct llama_opt_params lopt_params {
+        /*n_ctx_train     =*/ 0,
+        /*param_filter    =*/ llama_opt_param_filter_all,
+        /*param_filter_ud =*/ nullptr,
+        /*get_opt_pars    =*/ ggml_opt_get_constant_optimizer_params,
+        /*get_opt_pars_ud =*/ &optimizer_params,
+    };
+    llama_opt_init(ctx.get(), model.get(), lopt_params);
+
+    const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - val_split);
+
+    ggml_opt_result_t result_train = ggml_opt_result_init();
+    ggml_opt_result_t result_eval  = ggml_opt_result_init();
+
+    for (int epoch = 0; epoch < 2; ++epoch) {
+        llama_opt_epoch(ctx.get(), dataset, result_train, result_eval, idata_split,
+            ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
+        fprintf(stderr, "\n");
+
+        ggml_opt_result_reset(result_train);
+        ggml_opt_result_reset(result_eval);
+    }
+    ggml_opt_result_free(result_train);
+    ggml_opt_result_free(result_eval);
+
+    llama_model_save_to_file(model.get(), "finetuned-model.gguf");
+
+    llama_backend_free();
+
+    return 0;
+}

ggml/include/ggml-opt.h

@@ -37,13 +37,16 @@ extern "C" {
     // ====== Dataset ======
 
     GGML_API ggml_opt_dataset_t ggml_opt_dataset_init(
-            int64_t ne_datapoint, // number of elements per datapoint
-            int64_t ne_label,     // number of elements per label
-            int64_t ndata,        // total number of datapoints/labels
-            int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
+            enum ggml_type type_data,    // the type for the internal data tensor
+            enum ggml_type type_label,   // the type for the internal labels tensor
+            int64_t        ne_datapoint, // number of elements per datapoint
+            int64_t        ne_label,     // number of elements per label
+            int64_t        ndata,        // total number of datapoints/labels
+            int64_t        ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
     GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset);
 
     // get underlying tensors that store the data
+    GGML_API int64_t              ggml_opt_dataset_ndata (ggml_opt_dataset_t dataset);
     GGML_API struct ggml_tensor * ggml_opt_dataset_data  (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata]
     GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label,     ndata]
 
@@ -56,13 +59,19 @@ extern "C" {
             struct ggml_tensor * data_batch,   // shape = [ne_datapoint, ndata_batch]
             struct ggml_tensor * labels_batch, // shape = [ne_label,     ndata_batch]
             int64_t              ibatch);
+    GGML_API void ggml_opt_dataset_get_batch_host(
+            ggml_opt_dataset_t   dataset,
+            void               * data_batch,
+            size_t               nb_data_batch,
+            void               * labels_batch,
+            int64_t              ibatch);
 
     // ====== Model / Context ======
 
     enum ggml_opt_build_type {
-        GGML_OPT_BUILD_TYPE_FORWARD,
-        GGML_OPT_BUILD_TYPE_GRAD,
-        GGML_OPT_BUILD_TYPE_OPT,
+        GGML_OPT_BUILD_TYPE_FORWARD = 10,
+        GGML_OPT_BUILD_TYPE_GRAD    = 20,
+        GGML_OPT_BUILD_TYPE_OPT     = 30,
     };
 
     // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
@@ -81,20 +90,22 @@ extern "C" {
     // userdata can be used to pass arbitrary data
     typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata);
 
-    // returns the default optimizer params (constant)
+    // returns the default optimizer params (constant, hard-coded values)
     // userdata is not used
     GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
 
+    // casts userdata to ggml_opt_optimizer_params and returns it
+    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata);
+
     // parameters for initializing a new optimization context
     struct ggml_opt_params {
         ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs
 
-        struct ggml_context * ctx_compute; // created in user code, holds non-static tensors
-
-        // the forward graph is defined by inputs and outputs
-        // those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts
-        struct ggml_tensor * inputs;
-        struct ggml_tensor * outputs;
+        // by default the forward graph needs to be reconstructed for each eval
+        // if ctx_compute, inputs, and outputs are set the graphs are instead allocated statically
+        struct ggml_context * ctx_compute;
+        struct ggml_tensor  * inputs;
+        struct ggml_tensor  * outputs;
 
         enum ggml_opt_loss_type  loss_type;
         enum ggml_opt_build_type build_type;
@@ -107,12 +118,9 @@ extern "C" {
 
     // get parameters for an optimization context with defaults set where possible
     // parameters for which no sensible defaults exist are supplied as arguments to this function
-    GGML_API ggml_opt_params ggml_opt_default_params(
-            ggml_backend_sched_t      backend_sched,
-            struct ggml_context     * ctx_compute,
-            struct ggml_tensor      * inputs,
-            struct ggml_tensor      * outputs,
-            enum ggml_opt_loss_type   loss_type);
+    GGML_API struct ggml_opt_params ggml_opt_default_params(
+            ggml_backend_sched_t    backend_sched,
+            enum ggml_opt_loss_type loss_type);
 
     GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
     GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
@@ -121,6 +129,7 @@ extern "C" {
     GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
 
     // get underlying tensors that store data
+    // if not using static graphs these pointers become invalid with the next call to ggml_opt_alloc
     GGML_API struct ggml_tensor * ggml_opt_inputs(  ggml_opt_context_t opt_ctx); // forward graph input tensor
     GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor
     GGML_API struct ggml_tensor * ggml_opt_labels(  ggml_opt_context_t opt_ctx); // labels to compare outputs against
@@ -128,11 +137,12 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_opt_pred(    ggml_opt_context_t opt_ctx); // predictions made by outputs
     GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels
 
+    // get the gradient accumulator for a node from the forward graph
     GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
 
     // ====== Optimization Result ======
 
-    GGML_API ggml_opt_result_t ggml_opt_result_init();
+    GGML_API ggml_opt_result_t ggml_opt_result_init(void);
     GGML_API void ggml_opt_result_free(ggml_opt_result_t result);
     GGML_API void ggml_opt_result_reset(ggml_opt_result_t result);
 
@@ -144,11 +154,20 @@ extern "C" {
 
     // ====== Computation ======
 
-    // do forward pass, increment result if not NULL
-    GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+    // if not using static graphs, this function must be called prior to ggml_opt_alloc
+    GGML_API void ggml_opt_prepare_alloc(
+        ggml_opt_context_t    opt_ctx,
+        struct ggml_context * ctx_compute,
+        struct ggml_cgraph  * gf,
+        struct ggml_tensor  * inputs,
+        struct ggml_tensor  * outputs);
 
-    // do forward pass, increment result if not NULL, do backward pass
-    GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+    // allocate the next graph for evaluation, either forward or forward + backward
+    // must be called exactly once prior to calling ggml_opt_eval
+    GGML_API void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward);
+
+    // do forward pass, increment result if not NULL, do backward pass if allocated
+    GGML_API void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
 
     // ############################################################################
     // ## The high-level functions start here. They do not depend on any private ##
@@ -200,9 +219,9 @@ extern "C" {
     // fit model defined by inputs and outputs to dataset
     GGML_API void ggml_opt_fit(
             ggml_backend_sched_t            backend_sched,  // backend scheduler for constructing the compute graphs
-            ggml_context                  * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
-            ggml_tensor                   * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
-            ggml_tensor                   * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
+            struct ggml_context           * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
+            struct ggml_tensor            * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
+            struct ggml_tensor            * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
             ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
             enum ggml_opt_loss_type         loss_type,      // loss to minimize
             ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)

ggml/include/ggml.h

@@ -768,7 +768,7 @@ extern "C" {
     // Tensor flags
     GGML_API void ggml_set_input(struct ggml_tensor * tensor);
     GGML_API void ggml_set_output(struct ggml_tensor * tensor);
-    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API void ggml_set_param(struct ggml_tensor * tensor);
     GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
 
     //
@@ -938,7 +938,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_repeat_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b); // sum up values that are adjacent in dims > 0 instead of repeated with same stride
 
     // concat a and b along dim
     // used in stable-diffusion
@@ -2049,15 +2049,14 @@ extern "C" {
 
     GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
     GGML_API void ggml_build_backward_expand(
-        struct ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
-        struct ggml_context * ctx_compute, // context for gradient computation
-        struct ggml_cgraph  * cgraph,
-        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
+        struct ggml_context *  ctx,        // context for gradient computation
+        struct ggml_cgraph  *  cgraph,
+        struct ggml_tensor  ** grad_accs);
 
     // graph allocation in a context
     GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
     GGML_API struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads);
-    GGML_API struct ggml_cgraph * ggml_graph_dup       (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
+    GGML_API struct ggml_cgraph * ggml_graph_dup       (struct ggml_context * ctx, struct ggml_cgraph * cgraph, bool force_grads);
     GGML_API void                 ggml_graph_cpy       (struct ggml_cgraph * src, struct ggml_cgraph * dst);
     GGML_API void                 ggml_graph_reset     (struct ggml_cgraph * cgraph); // set regular grads + optimizer momenta to 0, set loss grad to 1
     GGML_API void                 ggml_graph_clear     (struct ggml_cgraph * cgraph);

ggml/src/ggml-backend.cpp

@@ -1111,7 +1111,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             const int node_backend_id = tensor_backend_id(node);
 
-            assert(node_backend_id != -1); // all nodes should be assigned by now
+            assert(node_backend_id != -1); // all nodes should be assigned by now, this can happen if there is no CPU fallback
 
             // check if we should start a new split based on the sources of the current node
             bool need_new_split = false;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -428,6 +428,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ${KLEIDIAI_SRC}/kai/ukernels/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
 
         set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
@@ -438,17 +439,19 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
         string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
 
-        set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS})
+        set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP})
 
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c)
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c)
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c)
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
+        list(APPEND GGML_KLEIDIAI_SOURCES
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
 
         if (NOT DOTPROD_ENABLED MATCHES -1)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
         endif()
 
         if (NOT I8MM_ENABLED MATCHES -1)
@@ -456,9 +459,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
 
         if (NOT SME_ENABLED MATCHES -1)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c)
-            set(PRIVATE_ARCH_FLAGS "${PRIVATE_ARCH_FLAGS}+sve+sve2")
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c)
+            set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
         endif()
 
         set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -4,16 +4,22 @@
 
 // KleidiAI micro-kernels
 #include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
-#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
-#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
-#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
-#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
+#include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
+
+#include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
+#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
+#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
+
+#include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h"
+#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
+#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
+
 #include "kai_common.h"
 
 #include "kernels.h"
@@ -61,6 +67,53 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+    {
+        /* SME GEMM */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+        },
+        /* SME GEMV */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .run_kernel            = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+        },
+        /* .lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme,
+            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme,
+            /* .pack_func             = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme,
+        },
+        /* .rhs_info = */ {
+            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .pack_func   = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_SME,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_F16,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #if defined(__APPLE__)
@@ -105,6 +158,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #if defined(__ARM_FEATURE_MATMUL_INT8)
@@ -148,6 +204,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #else
@@ -192,6 +251,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #if defined(__ARM_FEATURE_DOTPROD)
@@ -235,12 +297,33 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #endif
 };
 
-ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature features) {
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) {
+    ggml_kleidiai_kernels * kernel = nullptr;
+
+    if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
+        for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
+            if ((cpu_features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu &&
+                gemm_gemv_kernels[i].lhs_type == tensor->src[1]->type &&
+                gemm_gemv_kernels[i].rhs_type == tensor->src[0]->type &&
+                gemm_gemv_kernels[i].op_type  == tensor->type) {
+                kernel = &gemm_gemv_kernels[i];
+                break;
+            }
+        }
+    }
+
+    return kernel;
+}
+
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
     ggml_kleidiai_kernels * kernels = nullptr;
 
     for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {

ggml/src/ggml-cpu/kleidiai/kernels.h

@@ -4,6 +4,9 @@
 
 #pragma once
 
+#include <functional>
+#include "ggml.h"
+
 enum cpu_feature {
     CPU_FEATURE_NONE    = 0,
     CPU_FEATURE_DOTPROD = 1,
@@ -26,26 +29,53 @@ struct kernel_info {
     size_t (*get_nr)(void);
     size_t (*get_kr)(void);
     size_t (*get_sr)(void);
-    size_t (*get_lhs_offset)(size_t m_idx, size_t k, size_t bl);
-    size_t (*get_rhs_packed_offset)(size_t n_idx, size_t k, size_t bl);
+    std::variant<
+        std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
+        std::function<size_t(size_t m_idx, size_t k)>
+    > get_lhs_offset;
+    std::variant<
+        std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
+        std::function<size_t(size_t n_idx, size_t k)>
+    > get_rhs_packed_offset;
     size_t (*get_dst_offset)(size_t m_idx, size_t n_idx, size_t stride);
     size_t (*get_dst_size)(size_t m, size_t n);
-    void (*run_kernel)(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed,
-                         float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max);
+    std::variant<
+        std::function<void(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed,
+            float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max)>,
+        std::function<void(size_t m, size_t n, size_t k, const void* lhs_packed, const void* rhs_packed, void* dst, size_t dst_stride_row,
+            size_t dst_stride_col, float clamp_min, float clamp_max)>
+    > run_kernel;
 };
 
 struct lhs_packing_info {
     size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
-    size_t (*get_packed_offset)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
-    size_t (*packed_size)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
-    void (*pack_func)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
-                      size_t lhs_stride, void* lhs_packed);
+    std::variant<
+        std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
+        std::function<size_t(size_t m_idx, size_t k, size_t mr, size_t kr, size_t sr)>
+    > get_packed_offset;
+    std::variant<
+        std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
+        std::function<size_t(size_t m, size_t k, size_t mr, size_t kr, size_t sr)>
+    > packed_size;
+    std::variant<
+        std::function<void(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
+            size_t lhs_stride, void* lhs_packed)>,
+        std::function<void(size_t m, size_t k, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const void* lhs, size_t lhs_stride,
+        void* lhs_packed)>
+    > pack_func;
 };
 
 struct rhs_packing_info {
-    size_t (*packed_size)(size_t n, size_t k, size_t nr, size_t kr, size_t bl);
-    void (*pack_func)(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
-                      const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params);
+    std::variant<
+        std::function<size_t(size_t n, size_t k, size_t nr, size_t kr, size_t bl)>,
+        std::function<size_t(size_t n, size_t k)>
+    > packed_size;
+    std::variant<
+        std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
+            const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params)>,
+        std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride, const void* rhs,
+            const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params)>
+    > pack_func;
 };
 
 struct ggml_kleidiai_kernels {
@@ -55,6 +85,10 @@ struct ggml_kleidiai_kernels {
     rhs_packing_info rhs_info;
 
     cpu_feature required_cpu;
+    ggml_type lhs_type;
+    ggml_type rhs_type;
+    ggml_type op_type;
 };
 
-ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features);
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor);
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features);

ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -34,8 +34,9 @@
 #include "ggml-common.h"
 
 struct ggml_kleidiai_context {
+    cpu_feature features;
     ggml_kleidiai_kernels * kernels;
-} static ctx = { NULL };
+} static ctx = { CPU_FEATURE_NONE, NULL };
 
 static void init_kleidiai_context(void) {
 
@@ -47,18 +48,18 @@ static void init_kleidiai_context(void) {
         const char *env_var = getenv("GGML_KLEIDIAI_SME");
         int sme_enabled = 0;
 
-        cpu_feature features  = (ggml_cpu_has_dotprod()     ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
-                                (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
-                                (ggml_cpu_has_sve()         ? CPU_FEATURE_SVE     : CPU_FEATURE_NONE);
+        ctx.features  = (ggml_cpu_has_dotprod()     ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
+                        (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
+                        (ggml_cpu_has_sve()         ? CPU_FEATURE_SVE     : CPU_FEATURE_NONE);
 
         if (env_var) {
             sme_enabled = atoi(env_var);
         }
 
         if (sme_enabled != 0) {
-            features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
+            ctx.features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
         }
-        ctx.kernels = ggml_kleidiai_select_kernels(features);
+        ctx.kernels = ggml_kleidiai_select_kernels_q4_0(ctx.features);
     }
     ggml_critical_section_end();
 }
@@ -68,95 +69,275 @@ static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) {
     return tensor->ne[dim];
 }
 
+template<typename Ret, typename Variant, typename... Args>
+static Ret variant_call(const Variant & var, Args&&... args) {
+    return std::visit([&](auto&& func) -> Ret {
+        if constexpr (std::is_invocable_r_v<Ret, decltype(func), Args...>) {
+            return func(std::forward<Args>(args)...);
+        } else {
+            throw std::runtime_error("Invalid function type in variant_call");
+        }
+    }, var);
+}
+
 namespace ggml::cpu::kleidiai {
+
+static size_t round_down(size_t x, size_t y) {
+    return y == 0 ? x : x - (x % y);
+}
+
+static void transpose_f32kxn_f16nxk(size_t n, size_t k, float * dst, const uint16_t * src, size_t rhs_stride) {
+    size_t src_stride = rhs_stride / sizeof(uint16_t);
+    size_t dst_stride = n;
+
+    for (size_t k_idx = 0; k_idx < k; ++k_idx) {
+        for (size_t n_idx = 0; n_idx < n; ++n_idx) {
+            uint16_t v = *(src + k_idx + n_idx * src_stride);
+            *(dst + n_idx + k_idx * dst_stride) = kai_cast_f32_f16(v);
+        }
+    }
+}
+
 class tensor_traits : public ggml::cpu::tensor_traits {
     bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
-        GGML_ASSERT(ctx.kernels);
-        kernel_info * kernel = op->src[1]->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm;
+        ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
+        GGML_ASSERT(kernels);
+        kernel_info * kernel = op->src[1]->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
 
         size_t k = op->src[0]->ne[0];
+        size_t n = op->src[0]->ne[1];
         size_t m = op->src[1]->ne[1];
 
         size_t mr = kernel->get_mr();
         size_t kr = kernel->get_kr();
         size_t sr = kernel->get_sr();
 
-        size = ctx.kernels->lhs_info.packed_size(m, k, QK4_0, mr, kr, sr);
+        if (kernels->rhs_type == GGML_TYPE_Q4_0) {
+            size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, QK4_0, mr, kr, sr);
+        } else if (kernels->rhs_type == GGML_TYPE_F16) {
+            size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, mr, kr, sr) +
+                   variant_call<size_t>(kernels->rhs_info.packed_size, n, k) +
+                   k * n * sizeof(float) + n * sizeof(float);
+        } else {
+            GGML_ASSERT(false);
+        }
 
         return true;
     }
 
+
     bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * dst) override {
         if (dst->op == GGML_OP_MUL_MAT) {
-            const ggml_tensor * src0 = dst->src[0];
-            const ggml_tensor * src1 = dst->src[1];
+            if (dst->src[0]->type == GGML_TYPE_Q4_0) {
+                return compute_forward_q4_0(params, dst);
+            } else if (dst->src[0]->type == GGML_TYPE_F16) {
+                return compute_forward_kv_cache(params, dst);
+            }
+        }
+        return false;
+    }
 
-            GGML_TENSOR_BINARY_OP_LOCALS
+    bool compute_forward_kv_cache(ggml_compute_params * params, struct ggml_tensor * dst) {
+        static std::atomic_flag first_to_arrive = ATOMIC_FLAG_INIT;
 
-            GGML_ASSERT(ctx.kernels);
-            kernel_info * kernel = src1->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm;
-            lhs_packing_info * lhs_info = &ctx.kernels->lhs_info;
+        const ggml_tensor * src0 = dst->src[0];
+        const ggml_tensor * src1 = dst->src[1];
 
-            GGML_ASSERT(kernel);
+        GGML_TENSOR_BINARY_OP_LOCALS
 
-            const int ith = params->ith;
-            const int nth = params->nth;
+        ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
+        GGML_ASSERT(kernels);
 
-            const size_t k = ne00;
-            const size_t m = ne11;
-            const size_t n = ne01;
+        kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
+        GGML_ASSERT(kernel);
 
-            const size_t n_step = kernel->get_n_step();
-            const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
-            const size_t n_start = ith * num_n_per_thread;
+        const int nth = params->nth;
+        const int ith = params->ith;
 
-            size_t n_to_process = num_n_per_thread;
-            if ((n_start + n_to_process) > n) {
-                n_to_process = n - n_start;
+        const int64_t lhs_batch_size0 = ne12;
+        const int64_t rhs_batch_size0 = ne02;
+        const int64_t batch_size      = rhs_batch_size0;
+
+        const int64_t r = lhs_batch_size0 / rhs_batch_size0;
+
+        const int64_t m = ne11 * r;
+        const int64_t n = ne01;
+        const int64_t k = ne00;
+
+        const size_t lhs_stride = src1->nb[1];
+        const size_t rhs_stride = src0->nb[1];
+        const size_t dst_stride = dst->nb[1];
+
+        const int64_t mr = static_cast<int64_t>(kernel->get_mr());
+        const int64_t nr = static_cast<int64_t>(kernel->get_nr());
+        const int64_t kr = static_cast<int64_t>(kernel->get_kr());
+        const int64_t sr = static_cast<int64_t>(kernel->get_sr());
+
+        const size_t lhs_packed_size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, mr, kr, sr);
+        const size_t rhs_packed_size = variant_call<size_t>(kernels->rhs_info.packed_size, n, k);
+        const size_t kxn_size        = k * n * sizeof(float);
+        const size_t bias_size       = n * sizeof(float);
+
+        const size_t wsize_required = lhs_packed_size + rhs_packed_size + kxn_size + bias_size;
+        GGML_ASSERT(wsize_required <= params->wsize);
+
+        uint8_t * lhs_packed = static_cast<uint8_t *>(params->wdata);
+        uint8_t * rhs_packed = lhs_packed + lhs_packed_size;
+        uint8_t * rhs_kxn    = rhs_packed + rhs_packed_size;
+        uint8_t * bias       = rhs_kxn + kxn_size;
+
+        for (int64_t batch_idx = 0; batch_idx < batch_size; ++batch_idx) {
+            const uint8_t * lhs_batch = static_cast<const uint8_t *>(src1->data) + batch_idx * m * lhs_stride;
+            const uint8_t * rhs_batch = static_cast<const uint8_t *>(src0->data) + batch_idx * n * rhs_stride;
+            uint8_t * dst_batch       = static_cast<uint8_t *>(dst->data) + batch_idx * m * dst_stride;
+
+            // LHS packing
+            {
+                const int64_t m_roundup_mr = kai_roundup(m, mr);
+                const int64_t num_threads  = KAI_MIN(m_roundup_mr / mr, nth);
+
+                if (ith < num_threads) {
+                    const int64_t num_m_per_thread0   = round_down(m_roundup_mr / num_threads, mr);
+                    const int64_t num_m_per_threadN_1 = m - (num_threads - 1) * num_m_per_thread0;
+
+                    const int64_t m_start          = ith * num_m_per_thread0;
+                    const int64_t num_m_per_thread = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
+
+                    const size_t lhs_offset        = variant_call<size_t>(kernels->gemm.get_lhs_offset, m_start, lhs_stride);
+                    const size_t lhs_packed_offset = variant_call<size_t>(kernels->lhs_info.get_packed_offset, m_start, k, mr, kr, sr);
+
+                    const void * src_ptr = static_cast<const uint8_t *>(lhs_batch) + lhs_offset;
+                    void * dst_ptr       = static_cast<uint8_t *>(lhs_packed) + lhs_packed_offset;
+
+                    variant_call<void>(kernels->lhs_info.pack_func, num_m_per_thread, k, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr);
+                }
             }
 
-            const uint8_t * lhs        = static_cast<const uint8_t *>(src1->data);
-            uint8_t * lhs_packed       = (uint8_t*)params->wdata;
-            const uint8_t * rhs_packed = static_cast<const uint8_t *>(src0->data);
+            // RHS packing
+            if (first_to_arrive.test_and_set(std::memory_order_acquire) == false) {
+                // First thread to reach this point handles RHS packing
+                memset(bias, 0, n * sizeof(float));
+                transpose_f32kxn_f16nxk(n, k, reinterpret_cast<float *>(rhs_kxn),
+                                        reinterpret_cast<const uint16_t *>(rhs_batch), rhs_stride);
 
-            size_t mr = kernel->get_mr();
-            size_t kr = kernel->get_kr();
-            size_t sr = kernel->get_sr();
-
-            // Calculate number of columns to be processed per thread
-            const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
-            const size_t m_start = ith * num_m_per_thread;
-            size_t m_to_process = num_m_per_thread;
-            if ((m_start + m_to_process) > m) {
-                m_to_process = m - m_start;
-            }
-
-            if(m_start < m) {
-                // Transform LHS
-                const size_t src_stride        = src1->nb[1];
-                const float * src_ptr          = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1]));
-                const size_t lhs_packed_offset = lhs_info->get_packed_offset(m_start, k, QK4_0, mr, kr, sr);
-                void * lhs_packed_ptr          = static_cast<void *>(lhs_packed + lhs_packed_offset);
-
-                lhs_info->pack_func(m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
+                variant_call<void>(kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, n * sizeof(float),
+                             rhs_kxn, bias, nullptr, rhs_packed, 0, nullptr);
             }
 
             ggml_barrier(params->threadpool);
 
-            // Perform the operation
-            const size_t dst_stride        = dst->nb[1];
-            const size_t lhs_packed_offset = lhs_info->get_packed_offset(0, k, QK4_0, mr, kr, sr);
-            const size_t rhs_packed_offset = kernel->get_rhs_packed_offset(n_start, k, QK4_0);
-            const size_t dst_offset        = kernel->get_dst_offset(0, n_start, dst_stride);
-            const void * rhs_ptr           = static_cast<const void *>(rhs_packed + rhs_packed_offset);
-            const void* lhs_ptr            = (const void*)((const char *)lhs_packed + lhs_packed_offset);
-            float *dst_ptr                 = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
+            first_to_arrive.clear(std::memory_order_release);
 
-            kernel->run_kernel(m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr,
-                               dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
-            return true;
+            // Perform the matmul
+            {
+                const int64_t m_to_process = m;
+                const int64_t m_start      = 0;
+
+                const int64_t n_step      = static_cast<int64_t>(kernel->get_n_step());
+                const int64_t num_threads = KAI_MIN(n / n_step, nth);
+
+                if (ith < num_threads) {
+                    const int64_t num_n_per_thread0   = round_down(n / num_threads, n_step);
+                    const int64_t num_n_per_threadN_1 = n - (num_threads - 1) * num_n_per_thread0;
+
+                    const int64_t n_start      = ith * num_n_per_thread0;
+                    const int64_t n_to_process = (ith == num_threads - 1) ? num_n_per_threadN_1 : num_n_per_thread0;
+
+                    const size_t lhs_packed_offset = variant_call<size_t>(kernel->get_lhs_offset, m_start, k);
+                    const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, n_start, k);
+                    const size_t dst_offset        = kernel->get_dst_offset(m_start, n_start, dst_stride);
+
+                    const void * lhs_ptr = lhs_packed + lhs_packed_offset;
+                    const void * rhs_ptr = rhs_packed + rhs_packed_offset;
+                    float * dst_ptr      = reinterpret_cast<float *>(dst_batch + dst_offset);
+
+                    variant_call<void>(kernel->run_kernel, m_to_process, n_to_process, k, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
+                }
+            }
+
+            if (batch_idx != batch_size - 1) {
+                // This barrier is necessary when the batch size is larger than 1. While processing a batch,
+                // the work data buffer (params->wdata) is used as temporary storage which means that only
+                // a single batch can be processed at any given time. No barrier is needed for the last
+                // batch since GGML inserts a barrier between the execution of every operator.
+                ggml_barrier(params->threadpool);
+            }
         }
-        return false;
+
+        return true;
+    }
+
+    bool compute_forward_q4_0(struct ggml_compute_params * params, struct ggml_tensor * dst) {
+        const ggml_tensor * src0 = dst->src[0];
+        const ggml_tensor * src1 = dst->src[1];
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
+        GGML_ASSERT(kernels);
+
+        kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
+        lhs_packing_info * lhs_info = &kernels->lhs_info;
+
+        GGML_ASSERT(kernel);
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        const size_t k = ne00;
+        const size_t m = ne11;
+        const size_t n = ne01;
+
+        size_t mr = kernel->get_mr();
+        size_t kr = kernel->get_kr();
+        size_t sr = kernel->get_sr();
+
+        const uint8_t * lhs        = static_cast<const uint8_t *>(src1->data);
+        uint8_t * lhs_packed       = (uint8_t*)params->wdata;
+        const uint8_t * rhs_packed = static_cast<const uint8_t *>(src0->data);
+
+        const size_t n_step = kernel->get_n_step();
+        const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
+        const size_t n_start = ith * num_n_per_thread;
+
+        size_t n_to_process = num_n_per_thread;
+        if ((n_start + n_to_process) > n) {
+            n_to_process = n - n_start;
+        }
+
+        // Calculate number of columns to be processed per thread
+        const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
+        const size_t m_start = ith * num_m_per_thread;
+        size_t m_to_process = num_m_per_thread;
+        if ((m_start + m_to_process) > m) {
+            m_to_process = m - m_start;
+        }
+
+        if (m_start < m) {
+            // Transform LHS
+            const size_t src_stride        = src1->nb[1];
+            const float * src_ptr          = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1]));
+            const size_t lhs_packed_offset = variant_call<size_t>(lhs_info->get_packed_offset, m_start, k, QK4_0, mr, kr, sr);
+            void * lhs_packed_ptr          = static_cast<void *>(lhs_packed + lhs_packed_offset);
+
+            variant_call<void>(lhs_info->pack_func, m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        // Perform the operation
+        const size_t dst_stride        = dst->nb[1];
+        const size_t lhs_packed_offset = variant_call<size_t>(lhs_info->get_packed_offset, 0, k, QK4_0, mr, kr, sr);
+        const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, n_start, k, QK4_0);
+        const size_t dst_offset        = kernel->get_dst_offset(0, n_start, dst_stride);
+        const void * rhs_ptr           = static_cast<const void *>(rhs_packed + rhs_packed_offset);
+        const void* lhs_ptr            = (const void*)((const char *)lhs_packed + lhs_packed_offset);
+        float *dst_ptr                 = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
+
+        variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
+                           sizeof(float), -FLT_MAX, FLT_MAX);
+
+        return true;
     }
 
 public:
@@ -169,13 +350,13 @@ public:
         size_t sr      = ctx.kernels->gemm.get_sr();
 
 #ifndef NDEBUG
-        const size_t repacked_size = ctx.kernels->rhs_info.packed_size(n, k, nr, kr, QK4_0);
+        const size_t repacked_size = variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
         GGML_ASSERT(repacked_size <= data_size && "repacked size larger than the packed size!");
 #endif
         struct kai_rhs_pack_qs4cxs1s0_param params;
         params.lhs_zero_point = 1;
         params.rhs_zero_point = 8;
-        ctx.kernels->rhs_info.pack_func(1, n, k, nr, kr, sr, QK4_0, (const uint8_t *)data, NULL, tensor->data, 0, &params);
+        variant_call<void>(ctx.kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, QK4_0, (const uint8_t*)data, nullptr, tensor->data, 0, &params);
 
         return 0;
 
@@ -189,7 +370,7 @@ static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struc
 }
 }  // namespace ggml::cpu::kleidiai
 
-GGML_API enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+static enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor);
 
     GGML_UNUSED(buffer);
@@ -238,12 +419,11 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_b
 namespace ggml::cpu::kleidiai {
 class extra_buffer_type : ggml::cpu::extra_buffer_type {
     bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
-        if (    op->op == GGML_OP_MUL_MAT &&
-                op->src[0]->type == GGML_TYPE_Q4_0 &&
-                op->src[0]->buffer &&
-                (ggml_n_dims(op->src[0]) == 2) &&
-                op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels
-                ) {
+        if (op->op == GGML_OP_MUL_MAT &&
+            op->src[0]->type == GGML_TYPE_Q4_0 &&
+            op->src[0]->buffer &&
+            (ggml_n_dims(op->src[0]) == 2) &&
+            op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels) {
             if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
                 return false;
             }
@@ -260,6 +440,19 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
             if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
                 return (ggml::cpu::tensor_traits *) op->src[0]->extra;
             }
+            else if (ggml_kleidiai_select_kernels(ctx.features, op) &&
+                     op->src[0]->op == GGML_OP_VIEW &&
+                     (op->src[1]->op == GGML_OP_PERMUTE || op->src[1]->op ==  GGML_OP_SOFT_MAX) &&
+                     op->src[1]->ne[1] > 1) {
+                if ((op->src[0]->nb[0] != 2) ||
+                    (op->src[1]->nb[0] != 4) ||
+                    (op->src[0]->nb[1] * op->src[0]->ne[1] != op->src[0]->nb[2]) ||
+                    (op->src[1]->nb[1] * op->src[1]->ne[1] != op->src[1]->nb[2])) {
+                    return nullptr;
+                }
+
+                return ggml::cpu::kleidiai::get_tensor_traits(NULL, NULL);
+            }
         }
         return nullptr;
     }

ggml/src/ggml-cuda/acc.cu

@@ -1,47 +1,61 @@
 #include "acc.cuh"
 
-static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, int offset) {
-    const int i = blockDim.x * blockIdx.x + threadIdx.x;
+static __global__ void acc_f32(const float * x, const float * y, float * dst, const int64_t ne,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s11, const int64_t s12, const int64_t s13, const int64_t offset) {
+    const int64_t i = blockDim.x * blockIdx.x + threadIdx.x;
+
     if (i >= ne) {
         return;
     }
-    int src1_idx = i - offset;
-    int oz = src1_idx / nb2;
-    int oy = (src1_idx - (oz * nb2)) / nb1;
-    int ox = src1_idx % nb1;
-    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
-        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
-    } else {
-        dst[i] = x[i];
+
+    int64_t src1_idx = i - offset;
+
+    int64_t tmp = src1_idx;
+    const int64_t i13 = tmp / s13;
+    tmp -= i13 * s13;
+    const int64_t i12 = tmp / s12;
+    tmp -= i12 * s12;
+    const int64_t i11 = tmp / s11;
+    tmp -= i11 * s11;
+    const int64_t i10 = tmp;
+
+    float val = x[i];
+    if (src1_idx >= 0 && i10 < ne10 && i11 < ne11 && i12 < ne12 && i13 < ne13) {
+        val += y[((i13*ne12 + i12) * ne11 + i11) * ne10 + i10];
     }
+    dst[i] = val;
 }
 
-static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
-    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
-    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
+static void acc_f32_cuda(const float * x, const float * y, float * dst, const int64_t n_elements,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s1, const int64_t s2, const int64_t s3, const int64_t offset, cudaStream_t stream) {
+    const int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
+    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, ne13, s1, s2, s3, offset);
 }
 
 void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
-    const float * src0_d = (const float *)src0->data;
-    const float * src1_d = (const float *)src1->data;
-    float * dst_d = (float *)dst->data;
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       * dst_d  = (float       *)  dst->data;
+
     cudaStream_t stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
 
-    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
-    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
-    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
-    int offset = dst->op_params[3] / 4; // offset in bytes
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(dst->nb[0] == ggml_element_size(dst));
+    GGML_ASSERT(ggml_is_contiguously_allocated(dst));
 
-    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream);
+    const int64_t s1     = dst->op_params[0] / sizeof(float);
+    const int64_t s2     = dst->op_params[1] / sizeof(float);
+    const int64_t s3     = dst->op_params[2] / sizeof(float);
+    const int64_t offset = dst->op_params[3] / sizeof(float);
+
+    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], s1, s2, s3, offset, stream);
 }

ggml/src/ggml-cuda/sum.cu

@@ -31,7 +31,7 @@ void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguously_allocated(src0));
 
     const float * src0_d = (const float *) src0->data;
     float * dst_d = (float *) dst->data;

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -4855,8 +4855,6 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             if (!any_on_device) {
                 return false;
             }
-            GGML_ASSERT(ggml_is_contiguous(src0));
-            GGML_ASSERT(ggml_is_contiguous(src1));
             func = ggml_cl_add;
             break;
         case GGML_OP_MUL:

ggml/src/ggml-opt.cpp

@@ -28,16 +28,19 @@ struct ggml_opt_dataset {
 };
 
 struct ggml_opt_context {
-    ggml_backend_sched_t    backend_sched        = nullptr;
-    ggml_cgraph           * allocated_graph      = nullptr;
-    ggml_cgraph           * allocated_graph_copy = nullptr;
-    struct ggml_context   * ctx_static           = nullptr;
-    struct ggml_context   * ctx_static_cpu       = nullptr;
-    struct ggml_context   * ctx_compute          = nullptr;
-    struct ggml_context   * ctx_copy             = nullptr;
-    ggml_backend_buffer_t   buf_static           = nullptr;
-    ggml_backend_buffer_t   buf_static_cpu       = nullptr;
-    std::mt19937            rng;
+    ggml_backend_sched_t       backend_sched        = nullptr;
+    ggml_cgraph              * allocated_graph      = nullptr;
+    ggml_cgraph              * allocated_graph_copy = nullptr;
+    struct ggml_context      * ctx_static           = nullptr;
+    struct ggml_context      * ctx_cpu              = nullptr;
+    struct ggml_context      * ctx_compute          = nullptr;
+    struct ggml_context      * ctx_copy             = nullptr;
+    ggml_backend_buffer_t      buf_static           = nullptr;
+    ggml_backend_buffer_t      buf_cpu              = nullptr;
+    std::mt19937               rng;
+    enum ggml_opt_loss_type    loss_type;
+    enum ggml_opt_build_type   build_type;
+    enum ggml_opt_build_type   build_type_alloc;
 
     struct ggml_tensor * inputs  = nullptr;
     struct ggml_tensor * outputs = nullptr;
@@ -50,6 +53,11 @@ struct ggml_opt_context {
     struct ggml_cgraph * gf      = nullptr;
     struct ggml_cgraph * gb_grad = nullptr;
     struct ggml_cgraph * gb_opt  = nullptr;
+    bool static_graphs           = false;
+    bool eval_ready              = false;
+    std::vector<struct ggml_tensor *> grad_accs;
+    std::vector<struct ggml_tensor *> grad_m;
+    std::vector<struct ggml_tensor *> grad_v;
 
     int64_t iter               = 1;
     int32_t opt_period         = 1;
@@ -73,7 +81,13 @@ struct ggml_opt_result {
 
 // ====== Dataset ======
 
-ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, int64_t ndata, int64_t ndata_shard) {
+ggml_opt_dataset_t ggml_opt_dataset_init(
+        enum ggml_type type_data,
+        enum ggml_type type_label,
+        int64_t        ne_datapoint,
+        int64_t        ne_label,
+        int64_t        ndata,
+        int64_t        ndata_shard) {
     GGML_ASSERT(ne_datapoint >  0);
     GGML_ASSERT(ne_label     >= 0);
     GGML_ASSERT(ndata        >  0);
@@ -92,11 +106,11 @@ ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label,
         result->ctx = ggml_init(params);
     }
 
-    result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata);
+    result->data = ggml_new_tensor_2d(result->ctx, type_data, ne_datapoint, ndata);
     result->nbs_data = ggml_nbytes(result->data) * ndata_shard/ndata;
 
     if (ne_label > 0) {
-        result->labels = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_label, ndata);
+        result->labels = ggml_new_tensor_2d(result->ctx, type_label, ne_label, ndata);
         result->nbs_labels = ggml_nbytes(result->labels) * ndata_shard/ndata;
     } else {
         result->labels = nullptr;
@@ -119,6 +133,10 @@ void ggml_opt_dataset_free(ggml_opt_dataset_t dataset) {
     delete dataset;
 }
 
+int64_t ggml_opt_dataset_ndata(ggml_opt_dataset_t dataset) {
+    return dataset->ndata;
+}
+
 struct ggml_tensor * ggml_opt_dataset_data(ggml_opt_dataset_t dataset) {
     return dataset->data;
 }
@@ -144,6 +162,8 @@ void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor *
     GGML_ASSERT(   data_batch && ggml_is_contiguous(data_batch));
     GGML_ASSERT(!labels_batch || ggml_is_contiguous(labels_batch));
     GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
+    GGML_ASSERT(                   data_batch->type == dataset->data->type);
+    GGML_ASSERT(!labels_batch || labels_batch->type == dataset->labels->type);
 
     const size_t nb_data_batch = ggml_nbytes(data_batch);
     GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
@@ -171,6 +191,31 @@ void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor *
     }
 }
 
+void ggml_opt_dataset_get_batch_host(ggml_opt_dataset_t dataset, void * data_batch, size_t nb_data_batch, void * labels_batch, int64_t ibatch) {
+    GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
+    GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
+
+    const int64_t shards_per_batch = nb_data_batch / dataset->nbs_data;
+
+    GGML_ASSERT((ibatch + 1)*shards_per_batch <= int64_t(dataset->permutation.size()));
+
+    for (int64_t ishard_batch = 0; ishard_batch < shards_per_batch; ++ishard_batch) {
+        const int64_t ishard = dataset->permutation[ibatch*shards_per_batch + ishard_batch];
+
+        const char * ptr_data       = (const char *) dataset->data->data + ishard      *dataset->nbs_data;
+        char       * ptr_data_batch = (char       *) data_batch          + ishard_batch*dataset->nbs_data;
+        memcpy(ptr_data_batch, ptr_data, dataset->nbs_data);
+
+        if (!labels_batch) {
+            continue;
+        }
+
+        const char * ptr_labels       = (const char *) dataset->labels->data + ishard      *dataset->nbs_labels;
+        char       * ptr_labels_batch = (char       *) labels_batch          + ishard_batch*dataset->nbs_labels;
+        memcpy(ptr_labels_batch, ptr_labels, dataset->nbs_labels);
+    }
+}
+
 // ====== Model / Context ======
 
 struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata) {
@@ -187,17 +232,18 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us
     return result;
 }
 
+struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata) {
+    return *((struct ggml_opt_optimizer_params *) userdata);
+}
+
 struct ggml_opt_params ggml_opt_default_params(
         ggml_backend_sched_t      backend_sched,
-        struct ggml_context     * ctx_compute,
-        struct ggml_tensor      * inputs,
-        struct ggml_tensor      * outputs,
         enum ggml_opt_loss_type   loss_type) {
     return {
         /*backend_sched   =*/ backend_sched,
-        /*ctx_compute     =*/ ctx_compute,
-        /*inputs          =*/ inputs,
-        /*logits          =*/ outputs,
+        /*ctx_compute     =*/ nullptr,
+        /*inputs          =*/ nullptr,
+        /*logits          =*/ nullptr,
         /*loss_type       =*/ loss_type,
         /*build_type      =*/ GGML_OPT_BUILD_TYPE_OPT,
         /*opt_period      =*/ 1,
@@ -266,195 +312,246 @@ static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * src) {
     return dst;
 }
 
-static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) {
-    GGML_ASSERT(graph);
-    if (opt_ctx->allocated_graph == graph) {
-        return;
-    }
+static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
+    GGML_ASSERT(opt_ctx->ctx_compute && "no compute context set, either use static graphs or set one with ggml_opt_prepare_alloc");
+    GGML_ASSERT((!opt_ctx->static_graphs || opt_ctx->inputs->data) && "when using static graphs the inputs must be allocated statically");
 
-    ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
+    const bool accumulate = opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD &&
+        !(opt_ctx->static_graphs && opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period == 1);
 
-    {
-        ggml_init_params params = {
-            /*.mem_size   =*/ ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE,
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        ggml_free(opt_ctx->ctx_copy);
-        opt_ctx->ctx_copy = ggml_init(params);
-    }
-
-    opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
-
-    ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
-    opt_ctx->allocated_graph = graph;
-}
-
-ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
-    ggml_opt_context_t result = new struct ggml_opt_context;
-    result->backend_sched   = params.backend_sched;
-    result->ctx_compute     = params.ctx_compute;
-    result->inputs          = params.inputs;
-    result->outputs         = params.outputs;
-    result->opt_period      = params.opt_period;
-    result->get_opt_pars    = params.get_opt_pars;
-    result->get_opt_pars_ud = params.get_opt_pars_ud;
-
-    GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically");
-    GGML_ASSERT(result->opt_period >= 1);
-
-    const bool accumulate = params.build_type == GGML_OPT_BUILD_TYPE_GRAD ||
-        (params.build_type == GGML_OPT_BUILD_TYPE_OPT && result->opt_period > 1);
-
-    ggml_set_input(result->inputs);
-    ggml_set_output(result->outputs);
-
-    result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
-    ggml_build_forward_expand(result->gf, result->outputs);
+    ggml_set_input(opt_ctx->inputs);
+    ggml_set_output(opt_ctx->outputs);
 
     int n_param = 0;
-    for (int i = 0; i < result->gf->n_nodes; ++i) {
-        if (result->gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
+    for (int i = 0; i < opt_ctx->gf->n_nodes; ++i) {
+        const struct ggml_tensor * node = opt_ctx->gf->nodes[i];
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             n_param++;
         }
+        GGML_ASSERT(!(node->flags & GGML_TENSOR_FLAG_LOSS) && "support for extra loss terms not implemented");
     }
 
-    {
+    if (!opt_ctx->ctx_static) {
         // The static context is used for:
-        //   - gradients (1 tensor per param if using gradient accumulation)
+        //   - gradients (1 per loss, 1 tensor per param if using gradient accumulation)
         //   - optimizer momenta (2 tensors per param)
-        //   - labels
-        //   - loss + its gradient (up to 5 tensors)
-        //   - pred
-        //   - ncorrect (2 tensors).
-        const size_t tensors_per_param = (accumulate ? 1 : 0) + (params.build_type == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
-        const size_t size_meta = (tensors_per_param*n_param + 9) * ggml_tensor_overhead();
+        //   - labels (if using static graphs)
+        //   - loss (if using static graphs, up to 5 tensors)
+        //   - pred (if using static graphs)
+        //   - ncorrect (if using static graphs, 2 tensors).
+        constexpr size_t n_loss = 1;
+        const size_t tensors_per_param = (accumulate ? 1 : 0) +
+            (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
+        const size_t tensors_const = opt_ctx->static_graphs ? 9 : 0;
+        const size_t size_meta = (n_loss + tensors_per_param*n_param + tensors_const) * ggml_tensor_overhead();
         struct ggml_init_params params = {
             /*.mem_size   =*/ size_meta,
             /*.mem_buffer =*/ nullptr,
             /*.no_alloc   =*/ true,
         };
-        result->ctx_static = ggml_init(params);
+        opt_ctx->ctx_static = ggml_init(params);
     }
+    GGML_ASSERT(opt_ctx->build_type <= opt_ctx->build_type_alloc);
+
     {
-        // The static cpu context is used for:
-        //   - optimizer parameters (1 for the entire context)
+        // The cpu context is allocated statically if using static graphs, dynamically otherwise.
+        // It is used for:
+        //   - optimizer parameters (1 shared for all optimizer invocations)
         const size_t size_meta = 1 * ggml_tensor_overhead();
         struct ggml_init_params params = {
             /*.mem_size   =*/ size_meta,
             /*.mem_buffer =*/ nullptr,
             /*.no_alloc   =*/ true,
         };
-        result->ctx_static_cpu = ggml_init(params);
+        ggml_free(opt_ctx->ctx_cpu);
+        opt_ctx->ctx_cpu = ggml_init(params);
+
+        ggml_backend_buffer_free(opt_ctx->buf_cpu);
+        opt_ctx->buf_cpu = nullptr;
     }
 
+    struct ggml_context * ctx_results = opt_ctx->static_graphs ? opt_ctx->ctx_static : opt_ctx->ctx_compute;
 
-    switch (params.loss_type) {
+    switch (opt_ctx->loss_type) {
         case GGML_OPT_LOSS_TYPE_MEAN: {
-            result->loss = ggml_sum(result->ctx_static, result->outputs);
-            ggml_set_name(result->loss, "loss_sum");
-            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
-            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
-            ggml_set_name(result->loss, "loss_mean");
-            result->loss_per_datapoint = true;
+            opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->outputs);
+            ggml_set_name(opt_ctx->loss, "loss_sum");
+            const float scale = 1.0f / (opt_ctx->opt_period * ggml_nelements(opt_ctx->outputs));
+            opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, scale);
+            ggml_set_name(opt_ctx->loss, "loss_mean");
+            opt_ctx->loss_per_datapoint = true;
             break;
         }
         case GGML_OPT_LOSS_TYPE_SUM: {
-            result->loss = ggml_sum(result->ctx_static, result->outputs);
-            ggml_set_name(result->loss, "loss_sum");
-            result->loss_per_datapoint = false;
+            opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->outputs);
+            ggml_set_name(opt_ctx->loss, "loss_sum");
+            opt_ctx->loss_per_datapoint = false;
             break;
         }
         case GGML_OPT_LOSS_TYPE_CROSS_ENTROPY: {
-            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
-            ggml_set_input(result->labels);
-            ggml_set_name(result->labels, "labels");
-            result->loss = ggml_cross_entropy_loss(result->ctx_static, result->outputs, result->labels);
-            ggml_set_name(result->loss, "loss_cross_entropy");
-            if (result->opt_period > 1) {
-                result->loss = ggml_scale(result->ctx_static, result->loss, 1.0f / result->opt_period);
-                ggml_set_name(result->loss, "loss_cross_entropy_scaled");
+            opt_ctx->labels = ggml_dup_tensor(ctx_results, opt_ctx->outputs);
+            ggml_set_input(opt_ctx->labels);
+            ggml_set_name(opt_ctx->labels, "labels");
+            opt_ctx->loss = ggml_cross_entropy_loss(ctx_results, opt_ctx->outputs, opt_ctx->labels);
+            ggml_set_name(opt_ctx->loss, "loss_cross_entropy");
+            if (opt_ctx->opt_period > 1) {
+                opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, 1.0f / opt_ctx->opt_period);
+                ggml_set_name(opt_ctx->loss, "loss_cross_entropy_scaled");
             }
-            result->loss_per_datapoint = true;
+            opt_ctx->loss_per_datapoint = true;
             break;
         }
         case GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR: {
-            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
-            ggml_set_input(result->labels);
-            ggml_set_name(result->labels, "labels");
-            result->loss = ggml_sub(result->ctx_static, result->outputs, result->labels);
-            ggml_set_name(result->loss, "loss_error");
-            result->loss = ggml_sqr(result->ctx_static, result->loss);
-            ggml_set_name(result->loss, "loss_squared_error");
-            result->loss = ggml_sum(result->ctx_static, result->loss);
-            ggml_set_name(result->loss, "loss_sum_squared_error");
-            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
-            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
-            ggml_set_name(result->loss, "loss_mean_squared_error");
-            result->loss_per_datapoint = true;
+            opt_ctx->labels = ggml_dup_tensor(ctx_results, opt_ctx->outputs);
+            ggml_set_input(opt_ctx->labels);
+            ggml_set_name(opt_ctx->labels, "labels");
+            opt_ctx->loss = ggml_sub(ctx_results, opt_ctx->outputs, opt_ctx->labels);
+            ggml_set_name(opt_ctx->loss, "loss_error");
+            opt_ctx->loss = ggml_sqr(ctx_results, opt_ctx->loss);
+            ggml_set_name(opt_ctx->loss, "loss_squared_error");
+            opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->loss);
+            ggml_set_name(opt_ctx->loss, "loss_sum_squared_error");
+            const float scale = 1.0f / (opt_ctx->opt_period * ggml_nelements(opt_ctx->outputs));
+            opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, scale);
+            ggml_set_name(opt_ctx->loss, "loss_mean_squared_error");
+            opt_ctx->loss_per_datapoint = true;
             break;
         }
     }
-    ggml_set_output(result->loss);
-    ggml_set_loss(result->loss);
-    ggml_build_forward_expand(result->gf, result->loss);
+    ggml_set_output(opt_ctx->loss);
+    ggml_set_loss(opt_ctx->loss);
+    ggml_build_forward_expand(opt_ctx->gf, opt_ctx->loss);
 
-    result->pred = ggml_argmax(result->ctx_static, result->outputs);
-    ggml_set_name(result->pred, "pred");
-    ggml_set_output(result->pred);
-    ggml_build_forward_expand(result->gf, result->pred);
+    if (opt_ctx->loss_type == GGML_OPT_LOSS_TYPE_CROSS_ENTROPY) {
+        opt_ctx->pred = ggml_argmax(ctx_results, opt_ctx->outputs);
+        ggml_set_name(opt_ctx->pred, "pred");
+        ggml_set_output(opt_ctx->pred);
+        ggml_build_forward_expand(opt_ctx->gf, opt_ctx->pred);
 
-    if (result->labels) {
-        result->ncorrect = ggml_count_equal(result->ctx_static, result->pred, ggml_argmax(result->ctx_static, result->labels));
-        ggml_set_name(result->ncorrect, "ncorrect");
-        ggml_set_output(result->ncorrect);
-        ggml_build_forward_expand(result->gf, result->ncorrect);
-    } else {
-        result->ncorrect = nullptr;
+        opt_ctx->ncorrect = ggml_count_equal(ctx_results, opt_ctx->pred, ggml_argmax(ctx_results, opt_ctx->labels));
+        ggml_set_name(opt_ctx->ncorrect, "ncorrect");
+        ggml_set_output(opt_ctx->ncorrect);
+        ggml_build_forward_expand(opt_ctx->gf, opt_ctx->ncorrect);
     }
 
-    if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
-        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        return result;
+    if (opt_ctx->buf_static) {
+        if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
+            return;
+        }
+    } else if (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_FORWARD) {
+        opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(
+            opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
+        return;
     }
 
-    // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
-    result->gb_grad = ggml_graph_dup(result->ctx_compute, result->gf);
-    ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate);
+    if (opt_ctx->grad_accs.empty()) {
+        GGML_ASSERT(opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD);
 
-    if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) {
-        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        ggml_graph_reset(result->gb_grad);
-        return result;
-    }
+        const int n_nodes = opt_ctx->gf->n_nodes;
+        opt_ctx->grad_accs.resize(n_nodes);
+        for (int i = 0; i < n_nodes; ++i) {
+            ggml_tensor * node = opt_ctx->gf->nodes[i];
+            if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
+                opt_ctx->grad_accs[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+            } else {
+                opt_ctx->grad_accs[i] = nullptr;
+            }
+        }
 
-    GGML_ASSERT(params.build_type == GGML_OPT_BUILD_TYPE_OPT);
-
-    // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
-    result->gb_opt = ggml_graph_dup(result->ctx_compute, result->gb_grad);
-
-    result->adamw_params = ggml_new_tensor_1d(result->ctx_static_cpu, GGML_TYPE_F32, 7);
-    ggml_set_input(result->adamw_params);
-    ggml_set_name(result->adamw_params, "adamw_params");
-
-    for (int i = result->gf->n_nodes-1; i >= 0; --i) {
-        struct ggml_tensor * node = result->gb_opt->nodes[i];
-        struct ggml_tensor * grad = ggml_graph_get_grad(result->gb_opt, node);
-
-        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
-            struct ggml_tensor * m        = ggml_dup_tensor(result->ctx_static, node);
-            struct ggml_tensor * v        = ggml_dup_tensor(result->ctx_static, node);
-            struct ggml_tensor * opt_step = ggml_opt_step_adamw(result->ctx_compute, node, grad, m, v, result->adamw_params);
-            ggml_build_forward_expand(result->gb_opt, opt_step);
+        if (opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) {
+            opt_ctx->grad_m.resize(n_nodes);
+            opt_ctx->grad_v.resize(n_nodes);
+            for (int i = 0; i < n_nodes; ++i) {
+                ggml_tensor * node = opt_ctx->gf->nodes[i];
+                if (node->flags & GGML_TENSOR_FLAG_PARAM) {
+                    opt_ctx->grad_m[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+                    opt_ctx->grad_v[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+                } else {
+                    opt_ctx->grad_m[i] = nullptr;
+                    opt_ctx->grad_v[i] = nullptr;
+                }
+            }
         }
     }
 
-    result->buf_static = ggml_backend_alloc_ctx_tensors(
-        result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+    // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
+    opt_ctx->gb_grad = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gf, /*force_grads =*/ true);
+    ggml_build_backward_expand(opt_ctx->ctx_compute, opt_ctx->gb_grad, opt_ctx->grad_accs.data());
 
-    result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type());
+    if (opt_ctx->buf_static) {
+        if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_GRAD) {
+            return;
+        }
+    } else if (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_GRAD) {
+        opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
+        ggml_graph_reset(opt_ctx->gb_grad);
+    }
 
-    ggml_graph_reset(result->gb_opt);
+    GGML_ASSERT(opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT);
+
+    // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
+    opt_ctx->gb_opt = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gb_grad, /*force_grads =*/ true);
+
+    opt_ctx->adamw_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, 7);
+    ggml_set_input(opt_ctx->adamw_params);
+    ggml_set_name(opt_ctx->adamw_params, "adamw_params");
+
+    for (int i = opt_ctx->gf->n_nodes-1; i >= 0; --i) {
+        struct ggml_tensor * node = opt_ctx->gb_opt->nodes[i];
+        struct ggml_tensor * grad = ggml_graph_get_grad(opt_ctx->gb_opt, node);
+
+        if (grad && (node->flags & GGML_TENSOR_FLAG_PARAM)) {
+            struct ggml_tensor * m        = opt_ctx->grad_m[i];
+            struct ggml_tensor * v        = opt_ctx->grad_v[i];
+            struct ggml_tensor * opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, opt_ctx->adamw_params);
+
+            ggml_set_name(m,        (std::string("AdamW m for ")    + std::string(node->name)).c_str());
+            ggml_set_name(v,        (std::string("AdamW v for ")    + std::string(node->name)).c_str());
+            ggml_set_name(opt_step, (std::string("AdamW step for ") + std::string(node->name)).c_str());
+
+            ggml_build_forward_expand(opt_ctx->gb_opt, opt_step);
+        }
+    }
+
+    if (!opt_ctx->buf_static) {
+        opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(
+            opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
+        ggml_graph_reset(opt_ctx->gb_opt);
+    }
+
+    opt_ctx->buf_cpu = ggml_backend_alloc_ctx_tensors_from_buft(opt_ctx->ctx_cpu, ggml_backend_cpu_buffer_type());
+}
+
+ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
+    ggml_opt_context_t result = new struct ggml_opt_context;
+    result->backend_sched    = params.backend_sched;
+    result->ctx_compute      = params.ctx_compute;
+    result->loss_type        = params.loss_type;
+    result->build_type       = params.build_type;
+    result->build_type_alloc = params.build_type;
+    result->inputs           = params.inputs;
+    result->outputs          = params.outputs;
+    result->opt_period       = params.opt_period;
+    result->get_opt_pars     = params.get_opt_pars;
+    result->get_opt_pars_ud  = params.get_opt_pars_ud;
+
+    GGML_ASSERT(result->opt_period >= 1);
+
+    result->static_graphs = result->ctx_compute;
+
+    if (!result->static_graphs) {
+        GGML_ASSERT(!result->inputs);
+        GGML_ASSERT(!result->outputs);
+        return result;
+    }
+
+    GGML_ASSERT(result->inputs);
+    GGML_ASSERT(result->outputs);
+
+    result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
+    ggml_build_forward_expand(result->gf, result->outputs);
+
+    ggml_opt_build(result);
 
     return result;
 }
@@ -464,9 +561,9 @@ void ggml_opt_free(ggml_opt_context_t opt_ctx) {
         return;
     }
     ggml_backend_buffer_free(opt_ctx->buf_static);
-    ggml_backend_buffer_free(opt_ctx->buf_static_cpu);
+    ggml_backend_buffer_free(opt_ctx->buf_cpu);
     ggml_free(opt_ctx->ctx_static);
-    ggml_free(opt_ctx->ctx_static_cpu);
+    ggml_free(opt_ctx->ctx_cpu);
     delete opt_ctx;
 }
 
@@ -582,8 +679,79 @@ void ggml_opt_result_accuracy(ggml_opt_result_t result, double * accuracy, doubl
 
 // ====== Computation ======
 
-static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, ggml_opt_result * result) {
-    if (graph != opt_ctx->gf) {
+void ggml_opt_prepare_alloc(
+        ggml_opt_context_t    opt_ctx,
+        struct ggml_context * ctx_compute,
+        struct ggml_cgraph  * gf,
+        struct ggml_tensor  * inputs,
+        struct ggml_tensor  * outputs) {
+    GGML_ASSERT(!opt_ctx->static_graphs);
+    opt_ctx->ctx_compute = ctx_compute;
+    opt_ctx->gf          = gf;
+    opt_ctx->inputs      = inputs;
+    opt_ctx->outputs     = outputs;
+}
+
+void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward) {
+    GGML_ASSERT(!opt_ctx->eval_ready);
+    if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period > 1 && opt_ctx->opt_i == 0) {
+        ggml_graph_reset(opt_ctx->gb_grad);
+    }
+    if (backward) {
+        const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
+        opt_ctx->build_type = opt_i_next == 0 ? GGML_OPT_BUILD_TYPE_OPT : GGML_OPT_BUILD_TYPE_GRAD;
+    } else {
+        opt_ctx->build_type = GGML_OPT_BUILD_TYPE_FORWARD;
+    }
+
+    if (!opt_ctx->static_graphs) {
+        ggml_opt_build(opt_ctx);
+    }
+
+    struct ggml_cgraph * graph = nullptr;
+    switch (opt_ctx->build_type) {
+        case GGML_OPT_BUILD_TYPE_FORWARD: {
+            graph = opt_ctx->gf;
+        } break;
+        case GGML_OPT_BUILD_TYPE_GRAD: {
+            graph = opt_ctx->gb_grad;
+        } break;
+        case GGML_OPT_BUILD_TYPE_OPT: {
+            graph = opt_ctx->gb_opt;
+        } break;
+    }
+    GGML_ASSERT(graph);
+
+    if (opt_ctx->allocated_graph == graph) {
+        opt_ctx->eval_ready = true;
+        return;
+    }
+
+    ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
+
+    if (opt_ctx->static_graphs) {
+        ggml_init_params params = {
+            /*.mem_size   =*/ graph->size*ggml_tensor_overhead() + ggml_graph_overhead_custom(graph->size, graph->grads),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        ggml_free(opt_ctx->ctx_copy);
+        opt_ctx->ctx_copy = ggml_init(params);
+
+        opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
+    } else {
+        opt_ctx->allocated_graph_copy = graph;
+    }
+
+    ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
+    opt_ctx->allocated_graph = graph;
+
+    opt_ctx->eval_ready = true;
+}
+
+void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result) {
+    GGML_ASSERT(opt_ctx->eval_ready);
+    if (opt_ctx->allocated_graph == opt_ctx->gb_opt) {
         struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
 
         GGML_ASSERT(opt_pars.adamw.alpha >  0.0f);
@@ -609,9 +777,19 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph,
         adamw_par_data[6] = beta2h;
     }
 
-    ggml_opt_alloc_graph(opt_ctx, graph);
     ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
     opt_ctx->iter += opt_ctx->allocated_graph == opt_ctx->gb_opt;
+    opt_ctx->opt_i = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
+
+    if (!opt_ctx->static_graphs) {
+        opt_ctx->gf                   = nullptr;
+        opt_ctx->gb_grad              = nullptr;
+        opt_ctx->gb_opt               = nullptr;
+        opt_ctx->allocated_graph      = nullptr;
+        opt_ctx->allocated_graph_copy = nullptr;
+    }
+
+    opt_ctx->eval_ready = false;
 
     if (!result) {
         return;
@@ -635,12 +813,14 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph,
     ggml_backend_tensor_get(opt_ctx->loss, &loss, 0, ggml_nbytes(opt_ctx->loss));
     result->loss.push_back(loss);
 
-    GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
-    std::vector<int32_t> pred(ndata);
-    ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
-    result->pred.insert(result->pred.end(), pred.begin(), pred.end());
+    if (opt_ctx->pred) {
+        GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
+        std::vector<int32_t> pred(ndata);
+        ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
+        result->pred.insert(result->pred.end(), pred.begin(), pred.end());
+    }
 
-    if (!opt_ctx->labels || result->ncorrect < 0) {
+    if (!opt_ctx->ncorrect || result->ncorrect < 0) {
         result->ncorrect = -1;
         return;
     }
@@ -652,26 +832,6 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph,
     result->ncorrect += ncorrect;
 }
 
-void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
-    ggml_opt_eval_graph(opt_ctx, opt_ctx->gf, result);
-}
-
-void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
-    if (opt_ctx->opt_period == 1) {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
-        return;
-    }
-
-    const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
-    if (opt_i_next == 0) {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
-        ggml_opt_reset(opt_ctx, /*optimizer =*/ false);
-    } else {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_grad, result);
-    }
-    opt_ctx->opt_i = opt_i_next;
-}
-
 // ====== High-Level Functions ======
 
 void ggml_opt_epoch(
@@ -700,16 +860,18 @@ void ggml_opt_epoch(
     int64_t ibatch = 0;
     int64_t t_loop_start = ggml_time_us();
     for (; ibatch < ibatch_split; ++ibatch) {
+        ggml_opt_alloc(opt_ctx, /*backward =*/ true);
         ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
-        ggml_opt_forward_backward(opt_ctx, result_train);
+        ggml_opt_eval(opt_ctx, result_train);
         if (callback_train) {
             callback_train(true, opt_ctx, dataset, result_train, ibatch+1, ibatch_split, t_loop_start);
         }
     }
     t_loop_start = ggml_time_us();
     for (; ibatch < nbatches; ++ibatch) {
+        ggml_opt_alloc(opt_ctx, /*backward =*/ false);
         ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
-        ggml_opt_forward(opt_ctx, result_eval);
+        ggml_opt_eval(opt_ctx, result_eval);
         if (callback_eval) {
             callback_eval(false, opt_ctx, dataset, result_eval, ibatch+1-ibatch_split, nbatches-ibatch_split, t_loop_start);
         }
@@ -726,13 +888,26 @@ void ggml_opt_epoch_callback_progress_bar(
         int64_t            t_start_us) {
     fprintf(stderr, "%s[", train ? "train: " : "val:   ");
 
-    constexpr int64_t bar_length = 25;
+    // The progress bar consists of partially filled blocks, unicode has 8 separate fill levels.
+    constexpr int64_t bar_length = 8;
+    const int64_t ibatch8 = 8 * ibatch;
     for (int64_t j = 0; j < bar_length; ++j) {
-        const int64_t ibatch_j = ibatch_max * j/bar_length;
-        if (ibatch_j < ibatch) {
-            fprintf(stderr, "=");
-        } else if (ibatch_max * (j - 1)/bar_length < ibatch) {
-            fprintf(stderr, ">");
+        if        (ibatch_max * (8*j + 8) / bar_length < ibatch8) {
+            fprintf(stderr, "\u2588"); // full block
+        } else if (ibatch_max * (8*j + 7) / bar_length < ibatch8) {
+            fprintf(stderr, "\u2589"); // 7/8 filled
+        } else if (ibatch_max * (8*j + 6) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258A"); // 6/8 filled
+        } else if (ibatch_max * (8*j + 5) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258B"); // 5/8 filled
+        } else if (ibatch_max * (8*j + 4) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258C"); // 4/8 filled
+        } else if (ibatch_max * (8*j + 3) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258D"); // 3/8 filled
+        } else if (ibatch_max * (8*j + 2) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258E"); // 2/8 filled
+        } else if (ibatch_max * (8*j + 1) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258F"); // 1/8 filled
         } else {
             fprintf(stderr, " ");
         }
@@ -764,8 +939,8 @@ void ggml_opt_epoch_callback_progress_bar(
     const int64_t t_eta_m = t_eta_s / 60;
     t_eta_s -= t_eta_m * 60;
 
-    fprintf(stderr, "| data=%06" PRId64 "/%06" PRId64 ", loss=%.6lf+-%.6lf, accuracy=%.2lf+-%.2lf%%, "
-            "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 ", ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 "]\r",
+    fprintf(stderr, "] data=%07" PRId64 "/%07" PRId64 " loss=%.5lf±%.5lf acc=%.2lf±%.2lf%% "
+            "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 " ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 " \r",
             idata, idata_max, loss, loss_unc, 100.0*accuracy, 100.0*accuracy_unc,
             t_ibatch_h, t_ibatch_m, t_ibatch_s, t_eta_h, t_eta_m, t_eta_s);
     if (ibatch == ibatch_max) {
@@ -806,7 +981,10 @@ void ggml_opt_fit(
 
     int64_t epoch = 1;
 
-    ggml_opt_params params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type);
+    ggml_opt_params params = ggml_opt_default_params(backend_sched, loss_type);
+    params.ctx_compute     = ctx_compute;
+    params.inputs          = inputs;
+    params.outputs         = outputs;
     params.opt_period      = opt_period;
     params.get_opt_pars    = get_opt_pars;
     params.get_opt_pars_ud = &epoch;

ggml/src/ggml.c

@@ -5499,7 +5499,7 @@ static void ggml_compute_backward(
             // tensor = src0 * 1 + src1 * 0
             if (src0_needs_grads) {
                 // dsrc0 = dtensor * 1
-                ggml_add_or_set(ctx, cgraph, isrc0, grad);
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_reshape(ctx, grad, src0));
             }
             if (src1_needs_grads) {
                 // dsrc1 = dtensor * 0 -> noop
@@ -5780,10 +5780,9 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor *
 }
 
 void ggml_build_backward_expand(
-        struct ggml_context * ctx_static,
-        struct ggml_context * ctx_compute,
-        struct ggml_cgraph  * cgraph,
-        bool                  accumulate) {
+        struct ggml_context *  ctx,
+        struct ggml_cgraph  *  cgraph,
+        struct ggml_tensor  ** grad_accs) {
     GGML_ASSERT(cgraph->n_nodes > 0);
     GGML_ASSERT(cgraph->grads);
     GGML_ASSERT(cgraph->grad_accs);
@@ -5856,21 +5855,24 @@ void ggml_build_backward_expand(
         GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW ||
             node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE);
 
-        const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
-        GGML_ASSERT(igrad != GGML_HASHSET_FULL);
-        GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, igrad));
-        if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
-            cgraph->grad_accs[igrad] = ggml_dup_tensor(ctx_static, node);
-            cgraph->grads[igrad]     = cgraph->grad_accs[igrad];
-            ggml_format_name(cgraph->grad_accs[igrad], "grad acc for %s", node->name);
+        const size_t ihash = ggml_hash_find(&cgraph->visited_hash_set, node);
+        GGML_ASSERT(ihash != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, ihash));
+        if (grad_accs && grad_accs[i]) {
+            cgraph->grad_accs[ihash] = grad_accs[i];
+            cgraph->grads[ihash]     = cgraph->grad_accs[ihash];
+        } else if (node->flags & GGML_TENSOR_FLAG_LOSS) {
+            // loss tensors always need a gradient accumulator
+            cgraph->grad_accs[ihash] = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+            cgraph->grads[ihash]     = cgraph->grad_accs[ihash];
         }
-        grads_needed[igrad] = true;
+        grads_needed[ihash] = true;
     }
 
     for (int i = n_nodes_f - 1; i >= 0; --i) {
         // inplace operations to add gradients are not created by ggml_compute_backward except for gradient accumulation
         // use allocator to automatically make inplace operations
-        ggml_compute_backward(ctx_compute, cgraph, i, grads_needed);
+        ggml_compute_backward(ctx, cgraph, i, grads_needed);
     }
 
     free(grads_needed);
@@ -6016,8 +6018,8 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
     }
 }
 
-struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph) {
-    struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads != NULL);
+struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph, bool force_grads) {
+    struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads || force_grads);
     ggml_graph_cpy(cgraph, result);
     return result;
 }
@@ -6036,6 +6038,9 @@ struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) {
 }
 
 void ggml_graph_reset(struct ggml_cgraph * cgraph) {
+    if (!cgraph) {
+        return;
+    }
     GGML_ASSERT(cgraph->grads != NULL);
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
@@ -6345,8 +6350,8 @@ void ggml_set_output(struct ggml_tensor * tensor) {
     tensor->flags |= GGML_TENSOR_FLAG_OUTPUT;
 }
 
-void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor) {
-    GGML_UNUSED(ctx); // TODO: remove this parameter
+void ggml_set_param(struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->op == GGML_OP_NONE);
     tensor->flags |= GGML_TENSOR_FLAG_PARAM;
 }
 

gguf-py/gguf/constants.py

@@ -1905,6 +1905,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_GATE_EXP,
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
     ],
     MODEL_ARCH.CHAMELEON: [
         MODEL_TENSOR.TOKEN_EMBD,

gguf-py/gguf/tensor_mapping.py

@@ -428,6 +428,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.shared_expert.down_proj",  # qwen2moe
             "model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek deepseek2
             "language_model.model.layers.{bid}.feed_forward.shared_expert.down_proj", # llama4
+            "model.layers.{bid}.shared_mlp.output_linear",     # granitemoe
         ),
 
         MODEL_TENSOR.ATTN_Q_NORM: (

include/llama.h

@@ -4,6 +4,7 @@
 #include "ggml.h"
 #include "ggml-cpu.h"
 #include "ggml-backend.h"
+#include "ggml-opt.h"
 
 #include <stddef.h>
 #include <stdint.h>
@@ -344,7 +345,7 @@ extern "C" {
         float    yarn_beta_fast;   // YaRN low correction dim
         float    yarn_beta_slow;   // YaRN high correction dim
         uint32_t yarn_orig_ctx;    // YaRN original context size
-        float    defrag_thold;     // defragment the KV cache if holes/size > thold, < 0 disabled (default)
+        float    defrag_thold;     // defragment the KV cache if holes/size > thold, <= 0 disabled (default)
 
         ggml_backend_sched_eval_callback cb_eval;
         void * cb_eval_user_data;
@@ -445,6 +446,10 @@ extern "C" {
                                  size_t    n_paths,
               struct llama_model_params    params);
 
+    LLAMA_API void llama_model_save_to_file(
+            const struct llama_model * model,
+                        const char * path_model);
+
     DEPRECATED(LLAMA_API void llama_free_model(struct llama_model * model),
             "use llama_model_free instead");
 
@@ -1433,6 +1438,37 @@ extern "C" {
     LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
     LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
+    //
+    // training
+    //
+
+    // function that returns whether or not a given tensor contains trainable parameters
+    typedef bool (*llama_opt_param_filter)(const struct ggml_tensor * tensor, void * userdata);
+
+    // always returns true
+    LLAMA_API bool llama_opt_param_filter_all(const struct ggml_tensor * tensor, void * userdata);
+
+    struct llama_opt_params {
+        uint32_t n_ctx_train; // assumed context size post training, use context size specified in llama_context if 0
+
+        llama_opt_param_filter param_filter; // callback for determining which tensors contain trainable parameters
+        void * param_filter_ud;              // userdata for determining which tensors contain trainable parameters
+
+        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
+        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+    };
+
+    LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params);
+
+    LLAMA_API void llama_opt_epoch(
+            struct llama_context    * lctx,
+            ggml_opt_dataset_t        dataset,
+            ggml_opt_result_t         result_train,
+            ggml_opt_result_t         result_eval,
+            int64_t                   idata_split,
+            ggml_opt_epoch_callback   callback_train,
+            ggml_opt_epoch_callback   callback_eval);
+
 #ifdef __cplusplus
 }
 #endif

scripts/sync-ggml.last

@@ -1 +1 @@
-b59bddafe278877dfa22a80e53a637513862babb
+9b048bb72b811f50b0c30d9e5c84d6ff9f4bf005

src/CMakeLists.txt

@@ -23,6 +23,7 @@ add_library(llama
             llama-memory.cpp
             llama-mmap.cpp
             llama-model-loader.cpp
+            llama-model-saver.cpp
             llama-model.cpp
             llama-quant.cpp
             llama-sampling.cpp

src/llama-arch.cpp

@@ -1481,6 +1481,9 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
             { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,  "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
         },
     },
     {

src/llama-context.cpp

@@ -359,7 +359,9 @@ llama_context::llama_context(
     }
 }
 
-llama_context::~llama_context() = default;
+llama_context::~llama_context() {
+    ggml_opt_free(opt_ctx);
+}
 
 void llama_context::synchronize() {
     ggml_backend_sched_synchronize(sched.get());
@@ -1788,10 +1790,13 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
         }
     }
 
-    LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__);
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+    if (memory) {
+        LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__);
 
-    kv_self->state_read(io);
+        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+
+        kv_self->state_read(io);
+    }
 
     return io.n_bytes();
 }
@@ -1799,9 +1804,11 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
 size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) {
     GGML_UNUSED(seq_id);
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+    if (memory) {
+        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
 
-    kv_self->state_write(io, seq_id);
+        kv_self->state_write(io, seq_id);
+    }
 
     return io.n_bytes();
 }
@@ -1809,9 +1816,11 @@ size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id s
 size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) {
     GGML_UNUSED(seq_id);
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+    if (memory) {
+        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
 
-    kv_self->state_read(io, seq_id);
+        kv_self->state_read(io, seq_id);
+    }
 
     return io.n_bytes();
 }
@@ -1839,6 +1848,215 @@ void llama_context::perf_reset() {
     t_p_eval_us = n_p_eval = 0;
 }
 
+//
+// training
+//
+
+static void llama_set_param(struct ggml_tensor * tensor, llama_opt_param_filter param_filter, void * userdata) {
+    if (!tensor || tensor->type != GGML_TYPE_F32) {
+        return;
+    }
+    if (!param_filter(tensor, userdata)) {
+        return;
+    }
+    if (strcmp(tensor->name, "token_embd.weight") == 0) {
+        return; // FIXME
+    }
+    if (strcmp(tensor->name, "rope_freqs.weight") == 0) {
+        return; // FIXME
+    }
+    ggml_set_param(tensor);
+}
+
+void llama_context::opt_init(struct llama_model * model, struct llama_opt_params lopt_params) {
+    GGML_ASSERT(!opt_ctx);
+    model->hparams.n_ctx_train = lopt_params.n_ctx_train > 0 ? lopt_params.n_ctx_train : n_ctx();
+    const uint32_t n_batch     = std::min(this->n_batch(),  model->hparams.n_ctx_train);
+    const uint32_t n_ubatch    = std::min(this->n_ubatch(), n_batch);
+    GGML_ASSERT(model->hparams.n_ctx_train % n_batch  == 0);
+    GGML_ASSERT(n_batch                    % n_ubatch == 0);
+
+    ggml_opt_params opt_params = ggml_opt_default_params(sched.get(), GGML_OPT_LOSS_TYPE_CROSS_ENTROPY);
+    opt_params.opt_period      = n_batch / n_ubatch;
+    opt_params.get_opt_pars    = lopt_params.get_opt_pars;
+    opt_params.get_opt_pars_ud = lopt_params.get_opt_pars_ud;
+
+    opt_ctx = ggml_opt_init(opt_params);
+
+    llama_opt_param_filter param_filter = lopt_params.param_filter;
+    void * param_filter_ud              = lopt_params.param_filter_ud;
+
+  //llama_set_param(model->tok_embd,        param_filter, param_filter_ud); // FIXME
+    llama_set_param(model->type_embd,       param_filter, param_filter_ud);
+    llama_set_param(model->pos_embd,        param_filter, param_filter_ud);
+    llama_set_param(model->tok_norm,        param_filter, param_filter_ud);
+    llama_set_param(model->tok_norm_b,      param_filter, param_filter_ud);
+    llama_set_param(model->output_norm,     param_filter, param_filter_ud);
+    llama_set_param(model->output_norm_b,   param_filter, param_filter_ud);
+    llama_set_param(model->output,          param_filter, param_filter_ud);
+    llama_set_param(model->output_b,        param_filter, param_filter_ud);
+    llama_set_param(model->output_norm_enc, param_filter, param_filter_ud);
+    llama_set_param(model->cls,             param_filter, param_filter_ud);
+    llama_set_param(model->cls_b,           param_filter, param_filter_ud);
+    llama_set_param(model->cls_out,         param_filter, param_filter_ud);
+    llama_set_param(model->cls_out_b,       param_filter, param_filter_ud);
+
+    for (struct llama_layer & layer : model->layers) {
+        for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {
+            llama_set_param(reinterpret_cast<struct ggml_tensor **>(&layer)[i], param_filter, param_filter_ud);
+        }
+    }
+}
+
+void llama_context::opt_epoch_iter(
+        ggml_opt_dataset_t               dataset,
+        ggml_opt_result_t                result,
+        const std::vector<llama_token> & tokens,
+        const std::vector<llama_token> & labels_sparse,
+        llama_batch                    & batch,
+        ggml_opt_epoch_callback          callback,
+        bool                             train,
+        int64_t                          idata_in_loop,
+        int64_t                          ndata_in_loop,
+        int64_t                          t_loop_start) {
+    GGML_ASSERT(opt_ctx);
+    const uint32_t n_ctx    = llama_model_n_ctx_train(&model);
+    const uint32_t n_batch  = std::min(this->n_batch(),  n_ctx);
+    const uint32_t n_ubatch = std::min(this->n_ubatch(), n_batch);
+
+    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+
+    kv_self->clear();
+    llama_kv_cache_guard kv_guard(kv_self);
+
+    for (uint32_t pos_ctx = 0; pos_ctx < n_ctx; pos_ctx += n_batch) {
+        batch.n_tokens = n_batch;
+        for (uint32_t pos_batch = 0; pos_batch < n_batch; ++pos_batch) {
+            batch.token   [pos_batch]    = tokens[pos_ctx + pos_batch];
+            batch.pos     [pos_batch]    = pos_ctx + pos_batch;
+            batch.n_seq_id[pos_batch]    = 1;
+            batch.seq_id  [pos_batch][0] = 0;
+            batch.logits  [pos_batch]    = true;
+        }
+
+        const auto n_tokens_all = batch.n_tokens;
+
+        n_queued_tokens += n_tokens_all;
+
+        // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
+        const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
+
+        embd_seq.clear();
+
+        int64_t n_outputs_all = n_tokens_all;
+
+        llama_sbatch sbatch = kv_self->sbatch_init(batch, /*logits_all =*/ true);
+
+        // reserve output buffer
+        if (output_reserve(n_outputs_all) < n_outputs_all) {
+            LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all);
+            GGML_ABORT("TODO: handle this error");
+        };
+
+        for (uint32_t pos_batch = 0; pos_batch < n_batch; pos_batch += n_ubatch) {
+            llama_ubatch ubatch = kv_self->ubatch_next(sbatch, cparams.n_ubatch, embd_pooled);
+
+            n_outputs = ubatch.n_tokens;
+
+            // TODO: not sure if this is needed
+            if (!kv_self->find_slot(ubatch)) {
+                LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
+
+                GGML_ABORT("TODO: handle this error");
+            }
+
+            auto * gf = graph_init();
+            auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT);
+
+            struct ggml_context * ctx_compute_opt;
+            {
+                const size_t size_gf = ggml_graph_size(gf);
+                const size_t size_meta = 4*size_gf*ggml_tensor_overhead() + 2*ggml_graph_overhead_custom(size_gf, /*grads = */ true);
+                struct ggml_init_params params = {
+                    /*.mem_size   =*/ size_meta,
+                    /*.mem_buffer =*/ nullptr,
+                    /*.no_alloc   =*/ true,
+                };
+                ctx_compute_opt = ggml_init(params);
+            }
+            ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_tokens(), res->get_logits());
+            ggml_opt_alloc(opt_ctx, train);
+            res->set_inputs(&ubatch);
+            {
+                struct ggml_tensor * labels = ggml_opt_labels(opt_ctx);
+                GGML_ASSERT(labels->ne[1] == n_ubatch);
+                ggml_set_zero(labels);
+                const float onef = 1.0f;
+                for (uint32_t pos_ubatch = 0; pos_ubatch < n_ubatch; ++pos_ubatch) {
+                    const uint32_t ilabel = pos_ctx + pos_batch + pos_ubatch;
+                    GGML_ASSERT(labels_sparse[ilabel] < labels->ne[0]);
+                    ggml_backend_tensor_set(labels, &onef, (pos_ubatch*labels->ne[0] + labels_sparse[ilabel])*sizeof(float), sizeof(float));
+                }
+            }
+            ggml_opt_eval(opt_ctx, result);
+            if (callback) {
+                callback(train, opt_ctx, dataset, result, idata_in_loop + (pos_ctx + pos_batch)/n_ubatch + 1, ndata_in_loop, t_loop_start);
+            }
+            ggml_free(ctx_compute_opt);
+        }
+    }
+
+    kv_guard.commit();
+}
+
+void llama_context::opt_epoch(
+        ggml_opt_dataset_t        dataset,
+        ggml_opt_result_t         result_train,
+        ggml_opt_result_t         result_eval,
+        int64_t                   idata_split,
+        ggml_opt_epoch_callback   callback_train,
+        ggml_opt_epoch_callback   callback_eval) {
+    const uint32_t n_ctx    = this->n_ctx();
+    const uint32_t n_batch  = std::min(cparams.n_batch,  n_ctx);
+    const uint32_t n_ubatch = std::min(cparams.n_ubatch, n_batch);
+    const  int64_t ndata    = ggml_opt_dataset_ndata(dataset);
+
+    GGML_ASSERT(idata_split >= 0);
+    GGML_ASSERT(idata_split <= ndata);
+
+    const uint32_t ubatch_per_ctx = n_ctx / n_ubatch;
+
+    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
+    std::vector<llama_token>        tokens(n_ctx);
+    std::vector<llama_token> labels_sparse(n_ctx);
+
+    int64_t idata = 0;
+
+    int64_t t_loop_start = ggml_time_us();
+    int64_t ndata_in_loop = idata_split*ubatch_per_ctx;
+    for (; idata < idata_split; ++idata) {
+        constexpr bool train = true;
+        const int64_t idata_in_loop = idata*ubatch_per_ctx;
+
+        ggml_opt_dataset_get_batch_host(dataset, tokens.data(), n_ctx*sizeof(llama_token), labels_sparse.data(), idata);
+        opt_epoch_iter(dataset, result_train, tokens, labels_sparse, batch,
+            callback_train, train, idata_in_loop, ndata_in_loop, t_loop_start);
+    }
+
+    t_loop_start = ggml_time_us();
+    ndata_in_loop = (ndata - idata_split)*ubatch_per_ctx;
+    for (; idata < ndata; ++idata) {
+        constexpr bool train = false;
+        const int64_t idata_in_loop = (idata - idata_split)*ubatch_per_ctx;
+
+        ggml_opt_dataset_get_batch_host(dataset, tokens.data(), n_ctx*sizeof(llama_token), labels_sparse.data(), idata);
+        opt_epoch_iter(dataset, result_eval, tokens, labels_sparse, batch,
+            callback_eval, train, idata_in_loop, ndata_in_loop, t_loop_start);
+    }
+
+    llama_batch_free(batch);
+}
+
 //
 // interface implementation
 //
@@ -2457,3 +2675,34 @@ void llama_perf_context_print(const llama_context * ctx) {
 void llama_perf_context_reset(llama_context * ctx) {
     ctx->perf_reset();
 }
+
+//
+// training
+//
+
+bool llama_opt_param_filter_all(const struct ggml_tensor * tensor, void * userdata) {
+    GGML_UNUSED(tensor);
+    GGML_UNUSED(userdata);
+    return true;
+}
+
+void llama_opt_init(struct llama_context * ctx, struct llama_model * model, struct llama_opt_params lopt_params) {
+    ctx->opt_init(model, lopt_params);
+}
+
+void llama_opt_epoch(
+        struct llama_context    * ctx,
+        ggml_opt_dataset_t        dataset,
+        ggml_opt_result_t         result_train,
+        ggml_opt_result_t         result_eval,
+        int64_t                   idata_split,
+        ggml_opt_epoch_callback   callback_train,
+        ggml_opt_epoch_callback   callback_eval) {
+    ctx->opt_epoch(
+        dataset,
+        result_train,
+        result_eval,
+        idata_split,
+        callback_train,
+        callback_eval);
+}

src/llama-context.h

@@ -7,6 +7,7 @@
 #include "llama-adapter.h"
 
 #include "ggml-cpp.h"
+#include "ggml-opt.h"
 
 #include <map>
 #include <vector>
@@ -133,6 +134,32 @@ struct llama_context {
     llama_perf_context_data perf_get_data() const;
     void perf_reset();
 
+    //
+    // training
+    //
+
+    void opt_init(struct llama_model * model, struct llama_opt_params lopt_params);
+
+    void opt_epoch(
+            ggml_opt_dataset_t      dataset,
+            ggml_opt_result_t       result_train,
+            ggml_opt_result_t       result_eval,
+            int64_t                 idata_split,
+            ggml_opt_epoch_callback callback_train,
+            ggml_opt_epoch_callback callback_eval);
+
+    void opt_epoch_iter(
+            ggml_opt_dataset_t               dataset,
+            ggml_opt_result_t                result,
+            const std::vector<llama_token> & tokens,
+            const std::vector<llama_token> & labels_sparse,
+            llama_batch                    & batch,
+            ggml_opt_epoch_callback          callback,
+            bool                             train,
+            int64_t                          idata_in_loop,
+            int64_t                          ndata_in_loop,
+            int64_t                          t_loop_start);
+
 private:
     //
     // output
@@ -212,6 +239,9 @@ private:
 
     ggml_context_ptr ctx_compute;
 
+    // training
+    ggml_opt_context_t opt_ctx = nullptr;
+
     ggml_threadpool_t threadpool       = nullptr;
     ggml_threadpool_t threadpool_batch = nullptr;
 

src/llama-graph.cpp

@@ -971,6 +971,7 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
         inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
         //cb(inp->tokens, "inp_tokens", -1);
         ggml_set_input(inp->tokens);
+        res->t_tokens = inp->tokens;
 
         cur = ggml_get_rows(ctx0, tok_embd, inp->tokens);
 

src/llama-graph.h

@@ -298,6 +298,7 @@ class llm_graph_result_i {
 public:
     virtual ~llm_graph_result_i() = default;
 
+    virtual ggml_tensor * get_tokens()      = 0;
     virtual ggml_tensor * get_logits()      = 0;
     virtual ggml_tensor * get_embd()        = 0;
     virtual ggml_tensor * get_embd_pooled() = 0;
@@ -312,6 +313,7 @@ class llm_graph_result : public llm_graph_result_i {
 public:
     virtual ~llm_graph_result() = default;
 
+    ggml_tensor * get_tokens()      override { return t_tokens; }
     ggml_tensor * get_logits()      override { return t_logits; }
     ggml_tensor * get_embd()        override { return t_embd; }
     ggml_tensor * get_embd_pooled() override { return t_embd_pooled; }
@@ -328,6 +330,7 @@ public:
     }
 
     // important graph nodes
+    ggml_tensor * t_tokens      = nullptr;
     ggml_tensor * t_logits      = nullptr;
     ggml_tensor * t_embd        = nullptr;
     ggml_tensor * t_embd_pooled = nullptr;

src/llama-model-loader.cpp

@@ -301,12 +301,12 @@ namespace GGUFMeta {
             GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
 
         switch (arr_info.gt) {
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T, float>::value)); break;
-            case GGUF_TYPE_INT32:   GGML_ASSERT(
-                                            (std::is_same<T,  int32_t>::value) ||
-                                            (std::is_same<T, uint32_t>::value));  break;
+            case GGUF_TYPE_UINT32:
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
+                                                (std::is_same<T, uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
         }
 
         result.resize(arr_info.length);
@@ -330,12 +330,12 @@ namespace GGUFMeta {
             GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
 
         switch (arr_info.gt) {
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T, float>::value)); break;
-            case GGUF_TYPE_INT32:   GGML_ASSERT(
-                                            (std::is_same<T,  int32_t>::value) ||
-                                            (std::is_same<T, uint32_t>::value));  break;
+            case GGUF_TYPE_UINT32:
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
+                                                (std::is_same<T, uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
         }
 
         if (arr_info.length > N_MAX) {

src/llama-model-saver.cpp

@@ -0,0 +1,281 @@
+#include "llama-model-saver.h"
+
+#include "gguf.h"
+
+#include "llama.h"
+#include "llama-hparams.h"
+#include "llama-model.h"
+#include "llama-vocab.h"
+
+#include <string>
+
+llama_model_saver::llama_model_saver(const struct llama_model & model) : model(model), llm_kv(model.arch) {
+    gguf_ctx = gguf_init_empty();
+}
+
+llama_model_saver::~llama_model_saver() {
+    gguf_free(gguf_ctx);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const uint32_t value) {
+    gguf_set_val_u32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const int32_t value) {
+    gguf_set_val_i32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const float value) {
+    gguf_set_val_f32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const bool value) {
+    gguf_set_val_bool(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const char * value) {
+    gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+[[noreturn]]
+void llama_model_saver::add_kv(const enum llm_kv key, const char value) {
+    GGML_UNUSED(key);
+    GGML_UNUSED(value);
+    GGML_ABORT("fatal error"); // this should never be called, only needed to make the template below compile
+}
+
+template <typename Container>
+void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, const bool per_layer) {
+    const size_t n_values = per_layer ? size_t(model.hparams.n_layer) : value.size();
+    GGML_ASSERT(n_values <= value.size());
+
+    if (n_values == 0) {
+        return;
+    }
+
+    if (per_layer) {
+        bool all_values_the_same = true;
+        for (size_t i = 1; i < n_values; ++i) {
+            if (value[i] != value[0]) {
+                all_values_the_same = false;
+                break;
+            }
+        }
+        if (all_values_the_same) {
+            add_kv(key, value[0]);
+            return;
+        }
+    }
+
+    if (std::is_same<typename Container::value_type, uint8_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT8, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, int8_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT8, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, uint32_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT32, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, int32_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT32, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, float>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_FLOAT32, value.data(), n_values);
+    } else if (std::is_same<Container, std::string>::value) {
+        gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), reinterpret_cast<const char *>(value.data()));
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const std::vector<std::string> & value) {
+    std::vector<const char *> tmp(value.size());
+    for (size_t i = 0; i < value.size(); ++i) {
+        tmp[i] = value[i].c_str();
+    }
+    gguf_set_arr_str(gguf_ctx, llm_kv(key).c_str(), tmp.data(), tmp.size());
+}
+
+void llama_model_saver::add_tensor(const struct ggml_tensor * tensor) {
+    if (!tensor) {
+        return;
+    }
+    if (gguf_find_tensor(gguf_ctx, tensor->name) >= 0) {
+        GGML_ASSERT(std::string(tensor->name) == "rope_freqs.weight"); // FIXME
+        return;
+    }
+    gguf_add_tensor(gguf_ctx, tensor);
+}
+
+void llama_model_saver::add_kv_from_model() {
+    const llama_hparams & hparams = model.hparams;
+    const llama_vocab   & vocab   = model.vocab;
+
+    const int32_t n_vocab = vocab.n_tokens();
+    std::vector<std::string> tokens(n_vocab);
+    std::vector<float>       scores(n_vocab);
+    std::vector<int32_t>     token_types(n_vocab);
+
+    for (int32_t id = 0; id < n_vocab; ++id) {
+        const llama_vocab::token_data & token_data = vocab.get_token_data(id);
+
+        tokens[id] = token_data.text;
+        scores[id] = token_data.score;
+
+        switch(token_data.attr) {
+            case LLAMA_TOKEN_ATTR_UNKNOWN:      token_types[id] = LLAMA_TOKEN_TYPE_UNKNOWN;      break;
+            case LLAMA_TOKEN_ATTR_UNUSED:       token_types[id] = LLAMA_TOKEN_TYPE_UNUSED;       break;
+            case LLAMA_TOKEN_ATTR_NORMAL:       token_types[id] = LLAMA_TOKEN_TYPE_NORMAL;       break;
+            case LLAMA_TOKEN_ATTR_CONTROL:      token_types[id] = LLAMA_TOKEN_TYPE_CONTROL;      break;
+            case LLAMA_TOKEN_ATTR_USER_DEFINED: token_types[id] = LLAMA_TOKEN_TYPE_USER_DEFINED; break;
+            case LLAMA_TOKEN_ATTR_BYTE:         token_types[id] = LLAMA_TOKEN_TYPE_BYTE;         break;
+            case LLAMA_TOKEN_ATTR_UNDEFINED:
+            default:                            token_types[id] = LLAMA_TOKEN_TYPE_UNDEFINED;    break;
+        }
+    }
+
+    // add_kv(LLM_KV_GENERAL_TYPE,                      ???);
+    add_kv(LLM_KV_GENERAL_ARCHITECTURE,              model.arch_name());
+    // add_kv(LLM_KV_GENERAL_QUANTIZATION_VERSION,      ???);
+    // add_kv(LLM_KV_GENERAL_ALIGNMENT,                 ???);
+    add_kv(LLM_KV_GENERAL_NAME,                      model.name);
+    // add_kv(LLM_KV_GENERAL_AUTHOR,                    ???);
+    // add_kv(LLM_KV_GENERAL_VERSION,                   ???);
+    // add_kv(LLM_KV_GENERAL_URL,                       ???);
+    // add_kv(LLM_KV_GENERAL_DESCRIPTION,               ???);
+    // add_kv(LLM_KV_GENERAL_LICENSE,                   ???);
+    // add_kv(LLM_KV_GENERAL_SOURCE_URL,                ???);
+    // add_kv(LLM_KV_GENERAL_SOURCE_HF_REPO,            ???);
+
+    add_kv(LLM_KV_VOCAB_SIZE,                        vocab.n_tokens());
+    add_kv(LLM_KV_CONTEXT_LENGTH,                    hparams.n_ctx_train);
+    add_kv(LLM_KV_EMBEDDING_LENGTH,                  hparams.n_embd);
+    add_kv(LLM_KV_BLOCK_COUNT,                       hparams.n_layer);
+    add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);
+    add_kv(LLM_KV_FEED_FORWARD_LENGTH,               hparams.n_ff_arr, true);
+    add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    add_kv(LLM_KV_USE_PARALLEL_RESIDUAL,             hparams.use_par_res);
+    // add_kv(LLM_KV_TENSOR_DATA_LAYOUT,                ???);
+    add_kv(LLM_KV_EXPERT_COUNT,                      hparams.n_expert);
+    add_kv(LLM_KV_EXPERT_USED_COUNT,                 hparams.n_expert_used);
+    add_kv(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
+    add_kv(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale);
+    add_kv(LLM_KV_POOLING_TYPE,                      uint32_t(hparams.pooling_type));
+    add_kv(LLM_KV_LOGIT_SCALE,                       hparams.f_logit_scale);
+    add_kv(LLM_KV_DECODER_START_TOKEN_ID,            hparams.dec_start_token_id);
+    add_kv(LLM_KV_ATTN_LOGIT_SOFTCAPPING,            hparams.f_attn_logit_softcapping);
+    add_kv(LLM_KV_FINAL_LOGIT_SOFTCAPPING,           hparams.f_final_logit_softcapping);
+    add_kv(LLM_KV_SWIN_NORM,                         hparams.swin_norm);
+    add_kv(LLM_KV_RESCALE_EVERY_N_LAYERS,            hparams.rescale_every_n_layers);
+    add_kv(LLM_KV_TIME_MIX_EXTRA_DIM,                hparams.time_mix_extra_dim);
+    add_kv(LLM_KV_TIME_DECAY_EXTRA_DIM,              hparams.time_decay_extra_dim);
+    add_kv(LLM_KV_RESIDUAL_SCALE,                    hparams.f_residual_scale);
+    add_kv(LLM_KV_EMBEDDING_SCALE,                   hparams.f_embedding_scale);
+
+    add_kv(LLM_KV_ATTENTION_HEAD_COUNT,              hparams.n_head_arr, true);
+    add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV,           hparams.n_head_kv_arr, true);
+    add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS,          hparams.f_max_alibi_bias);
+    add_kv(LLM_KV_ATTENTION_CLAMP_KQV,               hparams.f_clamp_kqv);
+    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k);
+    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v);
+    add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS,           hparams.f_norm_eps);
+    add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
+    add_kv(LLM_KV_ATTENTION_CAUSAL,                  hparams.causal_attn);
+    add_kv(LLM_KV_ATTENTION_Q_LORA_RANK,             hparams.n_lora_q);
+    add_kv(LLM_KV_ATTENTION_KV_LORA_RANK,            hparams.n_lora_kv);
+    add_kv(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,  hparams.n_rel_attn_bkts);
+    add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW,          hparams.n_swa);
+    add_kv(LLM_KV_ATTENTION_SCALE,                   hparams.f_attention_scale);
+
+    const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
+
+    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot);
+    add_kv(LLM_KV_ROPE_FREQ_BASE,                    hparams.rope_freq_base_train);
+    // add_kv(LLM_KV_ROPE_SCALE_LINEAR,                 rope_scaling_factor); // old name
+    add_kv(LLM_KV_ROPE_SCALING_TYPE,                 llama_rope_scaling_type_name(hparams.rope_scaling_type_train));
+    add_kv(LLM_KV_ROPE_SCALING_FACTOR,               rope_scaling_factor);
+    add_kv(LLM_KV_ROPE_SCALING_ATTN_FACTOR,          hparams.rope_attn_factor);
+    add_kv(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,         hparams.n_ctx_orig_yarn);
+    add_kv(LLM_KV_ROPE_SCALING_FINETUNED,            hparams.rope_finetuned);
+    add_kv(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,         hparams.rope_yarn_log_mul);
+
+    // TODO: implement split file support
+    // add_kv(LLM_KV_SPLIT_NO,                          ???);
+    // add_kv(LLM_KV_SPLIT_COUNT,                       ???);
+    // add_kv(LLM_KV_SPLIT_TENSORS_COUNT,               ???);
+
+    add_kv(LLM_KV_SSM_INNER_SIZE,                    hparams.ssm_d_inner);
+    add_kv(LLM_KV_SSM_CONV_KERNEL,                   hparams.ssm_d_conv);
+    add_kv(LLM_KV_SSM_STATE_SIZE,                    hparams.ssm_d_state);
+    add_kv(LLM_KV_SSM_TIME_STEP_RANK,                hparams.ssm_dt_rank);
+    add_kv(LLM_KV_SSM_DT_B_C_RMS,                    hparams.ssm_dt_b_c_rms);
+
+    add_kv(LLM_KV_WKV_HEAD_SIZE,                     hparams.wkv_head_size);
+
+    add_kv(LLM_KV_TOKENIZER_MODEL,                   vocab.get_tokenizer_model());
+    add_kv(LLM_KV_TOKENIZER_PRE,                     vocab.get_tokenizer_pre());
+    add_kv(LLM_KV_TOKENIZER_LIST,                    tokens);
+    add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE,              token_types);
+    add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,        vocab.n_token_types());
+    add_kv(LLM_KV_TOKENIZER_SCORES,                  scores);
+    add_kv(LLM_KV_TOKENIZER_MERGES,                  vocab.get_bpe_merges());
+    // FIXME llama_token is type i32 but when reading in a GGUF file u32 is expected, not an issue for writing though
+    add_kv(LLM_KV_TOKENIZER_BOS_ID,                  uint32_t(vocab.token_bos()));
+    add_kv(LLM_KV_TOKENIZER_EOS_ID,                  uint32_t(vocab.token_eos()));
+    add_kv(LLM_KV_TOKENIZER_EOT_ID,                  uint32_t(vocab.token_eot()));
+    add_kv(LLM_KV_TOKENIZER_EOM_ID,                  uint32_t(vocab.token_eom()));
+    add_kv(LLM_KV_TOKENIZER_UNK_ID,                  uint32_t(vocab.token_unk()));
+    add_kv(LLM_KV_TOKENIZER_SEP_ID,                  uint32_t(vocab.token_sep()));
+    add_kv(LLM_KV_TOKENIZER_PAD_ID,                  uint32_t(vocab.token_pad()));
+    // add_kv(LLM_KV_TOKENIZER_CLS_ID,                  uint32_t(vocab.token_bos())); // deprecated
+    // add_kv(LLM_KV_TOKENIZER_MASK_ID,                 ???);
+    add_kv(LLM_KV_TOKENIZER_ADD_BOS,                 vocab.get_add_bos());
+    add_kv(LLM_KV_TOKENIZER_ADD_EOS,                 vocab.get_add_eos());
+    add_kv(LLM_KV_TOKENIZER_ADD_PREFIX,              vocab.get_add_space_prefix());
+    add_kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,         vocab.get_remove_extra_whitespaces());
+    add_kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,    vocab.get_precompiled_charsmap());
+    // add_kv(LLM_KV_TOKENIZER_HF_JSON,                 ???);
+    // add_kv(LLM_KV_TOKENIZER_RWKV,                    ???);
+    add_kv(LLM_KV_TOKENIZER_FIM_PRE_ID,              uint32_t(vocab.token_fim_pre()));
+    add_kv(LLM_KV_TOKENIZER_FIM_SUF_ID,              uint32_t(vocab.token_fim_suf()));
+    add_kv(LLM_KV_TOKENIZER_FIM_MID_ID,              uint32_t(vocab.token_fim_mid()));
+    add_kv(LLM_KV_TOKENIZER_FIM_PAD_ID,              uint32_t(vocab.token_fim_pad()));
+    add_kv(LLM_KV_TOKENIZER_FIM_REP_ID,              uint32_t(vocab.token_fim_rep()));
+    add_kv(LLM_KV_TOKENIZER_FIM_SEP_ID,              uint32_t(vocab.token_fim_sep()));
+
+    // TODO: implement LoRA support
+    // add_kv(LLM_KV_ADAPTER_TYPE,                      ???);
+    // add_kv(LLM_KV_ADAPTER_LORA_ALPHA,                ???);
+
+    // deprecated
+    // add_kv(LLM_KV_TOKENIZER_PREFIX_ID,               ???);
+    // add_kv(LLM_KV_TOKENIZER_SUFFIX_ID,               ???);
+    // add_kv(LLM_KV_TOKENIZER_MIDDLE_ID,               ???);
+}
+
+void llama_model_saver::add_tensors_from_model() {
+    if (std::string(model.output->name) != std::string(model.tok_embd->name)) {
+        add_tensor(model.tok_embd); // some models use the same tensor for tok_embd and output
+    }
+    add_tensor(model.type_embd);
+    add_tensor(model.pos_embd);
+    add_tensor(model.tok_norm);
+    add_tensor(model.tok_norm_b);
+    add_tensor(model.output_norm);
+    add_tensor(model.output_norm_b);
+    add_tensor(model.output);
+    add_tensor(model.output_b);
+    add_tensor(model.output_norm_enc);
+    add_tensor(model.cls);
+    add_tensor(model.cls_b);
+    add_tensor(model.cls_out);
+    add_tensor(model.cls_out_b);
+
+    for (const struct llama_layer & layer : model.layers) {
+        for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {
+            add_tensor(reinterpret_cast<const struct ggml_tensor * const *>(&layer)[i]);
+        }
+    }
+}
+
+void llama_model_saver::save(const std::string & path_model) {
+    gguf_write_to_file(gguf_ctx, path_model.c_str(), false);
+}
+

src/llama-model-saver.h

@@ -0,0 +1,37 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-arch.h"
+
+#include <vector>
+
+struct llama_model_saver {
+    struct gguf_context * gguf_ctx = nullptr;
+    const struct llama_model & model;
+    const struct LLM_KV llm_kv;
+
+    llama_model_saver(const struct llama_model & model);
+    ~llama_model_saver();
+
+    void add_kv(enum llm_kv key, uint32_t     value);
+    void add_kv(enum llm_kv key, int32_t      value);
+    void add_kv(enum llm_kv key, float        value);
+    void add_kv(enum llm_kv key, bool         value);
+    void add_kv(enum llm_kv key, const char * value);
+
+    [[noreturn]]
+    void add_kv(enum llm_kv key, char value); // needed to make the template below compile
+
+    template <typename Container>
+    void add_kv(enum llm_kv key, const Container & value, bool per_layer = false);
+
+    void add_kv(enum llm_kv key, const std::vector<std::string> & value);
+
+    void add_tensor(const struct ggml_tensor * tensor);
+
+    void add_kv_from_model();
+
+    void add_tensors_from_model();
+
+    void save(const std::string & path_model);
+};

src/llama-model.cpp

@@ -117,6 +117,10 @@ static const std::map<llama_rope_scaling_type, const char *> LLAMA_ROPE_SCALING_
     { LLAMA_ROPE_SCALING_TYPE_LONGROPE,   "longrope"   },
 };
 
+std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type) {
+    return LLAMA_ROPE_SCALING_TYPES.at(rope_scaling_type);
+}
+
 static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) {
     for (const auto & kv : LLAMA_ROPE_SCALING_TYPES) {
         if (kv.second == name) {
@@ -1385,6 +1389,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     // Add additional layer/vocab/etc checks here for other model sizes
                     default: type = LLM_TYPE_UNKNOWN;
                 }
+
+                // For Granite MoE Shared
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
             } break;
         case LLM_ARCH_CHAMELEON:
             {
@@ -1768,6 +1775,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
                             layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
                             layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+
+                            // For Granite MoE Shared
+                            if (hparams.n_ff_shexp > 0) {
+                                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
+                            }
                         }
                     }
                 } break;
@@ -4264,7 +4278,7 @@ uint64_t llama_model::n_elements() const {
 }
 
 void llama_model::print_info() const {
-    const char * rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train);
+    const std::string rope_scaling_type = llama_rope_scaling_type_name(hparams.rope_scaling_type_train);
 
     auto print_f = [](const std::function<uint32_t(uint32_t)> & f, uint32_t n) {
         bool is_var = false;
@@ -4325,7 +4339,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: causal attn      = %d\n",     __func__, hparams.causal_attn);
         LLAMA_LOG_INFO("%s: pooling type     = %d\n",     __func__, hparams.pooling_type);
         LLAMA_LOG_INFO("%s: rope type        = %d\n",     __func__, hparams.rope_type);
-        LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type);
+        LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type.c_str());
         LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
@@ -4381,10 +4395,13 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
     }
 
-    if (arch == LLM_ARCH_MINICPM || arch == LLM_ARCH_GRANITE || arch == LLM_ARCH_GRANITE_MOE) {
+    if (arch == LLM_ARCH_MINICPM ||
+        arch == LLM_ARCH_GRANITE ||
+        arch == LLM_ARCH_GRANITE_MOE) {
         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
+        LLAMA_LOG_INFO("%s: n_ff_shexp        = %d\n", __func__, hparams.n_ff_shexp);
     }
 
     if (arch == LLM_ARCH_BAILINGMOE) {
@@ -4594,11 +4611,6 @@ struct llm_build_llama : public llm_graph_context {
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
 
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
             ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
             cb(ffn_inp, "ffn_inp", il);
 
@@ -4670,11 +4682,6 @@ struct llm_build_llama : public llm_graph_context {
                 cb(cur, "ffn_moe_out", il);
             }
 
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
             cur = ggml_add(ctx0, cur, ffn_inp);
             cb(cur, "ffn_out", il);
 
@@ -4697,11 +4704,6 @@ struct llm_build_llama : public llm_graph_context {
         // lm_head
         cur = build_lora_mm(model.output, cur);
 
-        // For Granite architecture
-        if (hparams.f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
-        }
-
         cb(cur, "result_output", -1);
         res->t_logits = cur;
 
@@ -4812,11 +4814,6 @@ struct llm_build_deci : public llm_graph_context {
                 continue;
             }
 
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
             // modified to support attention-free layer of Llama-3_1-Nemotron-51B
             ggml_tensor * ffn_inp = cur;
             if (n_head > 0) {
@@ -4840,11 +4837,6 @@ struct llm_build_deci : public llm_graph_context {
                 cb(cur, "ffn_out", il);
             }
 
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
             cur = ggml_add(ctx0, cur, ffn_inp);
             cb(cur, "ffn_out", il);
 
@@ -4867,11 +4859,6 @@ struct llm_build_deci : public llm_graph_context {
         // lm_head
         cur = build_lora_mm(model.output, cur);
 
-        // For Granite architecture
-        if (hparams.f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
-        }
-
         cb(cur, "result_output", -1);
         res->t_logits = cur;
 
@@ -12210,6 +12197,195 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base {
     }
 };
 
+
+struct llm_build_granite : public llm_graph_context {
+    llm_build_granite(
+        const llama_model & model,
+        const llm_graph_params & params,
+        ggml_cgraph * gf,
+        const bool use_rope = true)
+        : llm_graph_context(params) {
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - built only if rope enabled
+        ggml_tensor * inp_pos = nullptr;
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and (optionally) RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                if (use_rope) {
+
+                    if (!inp_pos) {
+                        inp_pos = build_inp_pos();
+                    }
+                    ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                    Qcur = ggml_rope_ext(
+                            ctx0, Qcur, inp_pos, rope_factors,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+
+                    Kcur = ggml_rope_ext(
+                            ctx0, Kcur, inp_pos, rope_factors,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                }
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // For Granite architectures - scale residual
+            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network (non-MoE)
+            if (model.layers[il].ffn_gate_inp == nullptr) {
+
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+
+            } else {
+                // MoE branch
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                ggml_tensor * moe_out = build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, true,
+                        false, 0.0,
+                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                        il);
+                cb(moe_out, "ffn_moe_out", il);
+
+                // For Granite MoE Shared
+                if (hparams.n_ff_shexp > 0) {
+                    ggml_tensor * ffn_shexp = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                    cb(ffn_shexp, "ffn_shexp", il);
+
+                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                    cb(cur, "ffn_out", il);
+                } else {
+                    cur = moe_out;
+                }
+            }
+
+            // For Granite architectures - scale residual
+            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        // For Granite architectures - scale logits
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 // ref: https://github.com/facebookresearch/chameleon
 // based on the original build_llama() function, changes:
 //   * qk-norm
@@ -12917,8 +13093,6 @@ llm_graph_result_ptr llama_model::build_graph(
         case LLM_ARCH_LLAMA:
         case LLM_ARCH_LLAMA4:
         case LLM_ARCH_MINICPM:
-        case LLM_ARCH_GRANITE:
-        case LLM_ARCH_GRANITE_MOE:
             {
                 llm = std::make_unique<llm_build_llama>(*this, params, gf);
             } break;
@@ -13149,6 +13323,11 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_arwkv7>(*this, params, gf);
             } break;
+        case LLM_ARCH_GRANITE:
+        case LLM_ARCH_GRANITE_MOE:
+            {
+                llm = std::make_unique<llm_build_granite>(*this, params, gf);
+            } break;
         case LLM_ARCH_CHAMELEON:
             {
                 llm = std::make_unique<llm_build_chameleon>(*this, params, gf);

src/llama-model.h

@@ -96,6 +96,8 @@ enum llm_type {
     LLM_TYPE_235B_A22B,
 };
 
+std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type);
+
 struct llama_layer_posnet {
     // resnet
     struct ggml_tensor * norm1   = nullptr;

src/llama-quant.cpp

@@ -519,7 +519,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         nthread = std::thread::hardware_concurrency();
     }
 
-    // mmap consistently increases speed Linux, and also increases speed on Windows with
+    // mmap consistently increases speed on Linux, and also increases speed on Windows with
     // hot cache. It may cause a slowdown on macOS, possibly related to free memory.
 #if defined(__linux__) || defined(_WIN32)
     constexpr bool use_mmap = true;
@@ -529,7 +529,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
     llama_model_kv_override * kv_overrides = nullptr;
     if (params->kv_overrides) {
-        auto v = (std::vector<llama_model_kv_override>*)params->kv_overrides;
+        auto * v = (std::vector<llama_model_kv_override>*)params->kv_overrides;
         kv_overrides = v->data();
     }
 

src/llama-vocab.cpp

@@ -1,5 +1,7 @@
 #include "llama-vocab.h"
 
+#include "ggml.h"
+#include "gguf.h"
 #include "llama-impl.h"
 #include "llama-model-loader.h"
 
@@ -1234,6 +1236,9 @@ struct fragment_buffer_variant {
 struct llama_vocab::impl {
     uint32_t n_token_types = 0; // for BERT-style token types
 
+    std::string tokenizer_model;
+    std::string tokenizer_pre;
+
     enum llama_vocab_type     type     = LLAMA_VOCAB_TYPE_SPM;
     enum llama_vocab_pre_type pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
 
@@ -1369,9 +1374,6 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
     // determine vocab type
     {
-        std::string tokenizer_model;
-        std::string tokenizer_pre;
-
         ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_model);
         ml.get_key(LLM_KV_TOKENIZER_PRE,   tokenizer_pre, false);
 
@@ -1466,7 +1468,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
             const int precompiled_charsmap_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP).c_str());
             if (precompiled_charsmap_keyidx != -1) {
-                size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx);
+                const gguf_type pc_type = gguf_get_arr_type(ctx, precompiled_charsmap_keyidx);
+                GGML_ASSERT(pc_type == GGUF_TYPE_INT8 || pc_type == GGUF_TYPE_UINT8);
+
+                const size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx);
                 const char * pc = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx);
                 precompiled_charsmap.assign(pc, pc + n_precompiled_charsmap);
 #ifdef IS_BIG_ENDIAN
@@ -2789,6 +2794,14 @@ void llama_vocab::load(llama_model_loader & ml, const LLM_KV & kv) {
     pimpl->load(ml, kv);
 }
 
+std::string llama_vocab::get_tokenizer_model() const {
+    return pimpl->tokenizer_model;
+}
+
+std::string llama_vocab::get_tokenizer_pre() const {
+    return pimpl->tokenizer_pre;
+}
+
 enum llama_vocab_type llama_vocab::get_type() const {
     return pimpl->type;
 }
@@ -3011,6 +3024,20 @@ int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string
     return it->second;
 }
 
+std::vector<std::string> llama_vocab::get_bpe_merges() const {
+    std::vector<std::string> result(pimpl->bpe_ranks.size());
+
+    for (const auto & pair : pimpl->bpe_ranks) {
+        result[pair.second] = pair.first.first + " " + pair.first.second;
+    }
+
+    return result;
+}
+
+std::vector<char> llama_vocab::get_precompiled_charsmap() const {
+    return pimpl->precompiled_charsmap;
+}
+
 int32_t llama_vocab::tokenize(
                   const char * text,
                      int32_t   text_len,

src/llama-vocab.h

@@ -21,6 +21,9 @@ struct llama_vocab {
 
     void load(llama_model_loader & ml, const LLM_KV & kv);
 
+    std::string get_tokenizer_model() const;
+    std::string get_tokenizer_pre() const;
+
     enum llama_vocab_type     get_type()     const;
     enum llama_vocab_pre_type get_pre_type() const;
 
@@ -80,6 +83,9 @@ struct llama_vocab {
     int max_token_len() const;
 
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const;
+    std::vector<std::string> get_bpe_merges() const;
+
+    std::vector<char> get_precompiled_charsmap() const;
 
     int32_t tokenize(
                    const char * text,

src/llama.cpp

@@ -4,6 +4,7 @@
 #include "llama-mmap.h"
 #include "llama-vocab.h"
 #include "llama-model-loader.h"
+#include "llama-model-saver.h"
 #include "llama-model.h"
 
 #include "ggml.h"
@@ -253,6 +254,13 @@ struct llama_model * llama_model_load_from_splits(
     return llama_model_load_from_file_impl(splits.front(), splits, params);
 }
 
+void llama_model_save_to_file(const struct llama_model * model, const char * path_model) {
+    llama_model_saver ms(*model);
+    ms.add_kv_from_model();
+    ms.add_tensors_from_model();
+    ms.save(path_model);
+}
+
 //
 // chat templates
 //
@@ -338,3 +346,4 @@ const char * llama_print_system_info(void) {
 
     return s.c_str();
 }
+

tests/test-backend-ops.cpp

@@ -823,7 +823,7 @@ struct test_case {
 
         ggml_build_forward_expand(gf, out);
         ggml_graph_cpy(gf, gb);
-        ggml_build_backward_expand(ctx.get(), ctx.get(), gb, false);
+        ggml_build_backward_expand(ctx.get(), gb, nullptr);
         if (expect.size() != 1 || expect[0] != 0.0f) {
             GGML_ASSERT(ggml_graph_n_nodes(gb) > ggml_graph_n_nodes(gf));
             for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != NULL; t = ggml_get_next_tensor(ctx.get(), t)) {
@@ -1026,7 +1026,7 @@ struct test_example : public test_case {
         // Step 3: return the output tensor.
         return out;
     }
-    // In order to also check the gradients for your op, add calls like ggml_set_param(ctx, a)
+    // In order to also check the gradients for your op, add calls like ggml_set_param(a)
     // immediately after you create the tensors.
     // This is optional and only makes sense if a backward pass has actually been implemented for the new op.
 };
@@ -1058,7 +1058,7 @@ struct test_unary : public test_case {
             auto ne = ne_a; ne[0] *= 3;
             a = ggml_new_tensor(ctx, type, 4, ne.data());
             if (grad_supported) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
 
@@ -1067,7 +1067,7 @@ struct test_unary : public test_case {
         } else {
             a = ggml_new_tensor(ctx, type, 4, ne_a.data());
             if (grad_supported) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
         }
@@ -1133,7 +1133,7 @@ struct test_get_rows : public test_case {
 
         const bool grad_supported = ggml_is_matrix(in) && ggml_is_vector(rows);
         if (grad_supported) {
-            ggml_set_param(ctx, in);
+            ggml_set_param(in);
             // rows is a constant input -> no gradients
         }
 
@@ -1322,7 +1322,7 @@ struct test_repeat : public test_case {
         ggml_set_name(target, "target");
 
         ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
 
         ggml_tensor * out = ggml_repeat(ctx, src, target);
@@ -1406,7 +1406,7 @@ struct test_dup : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
 
         if (_use_permute) {
@@ -1442,7 +1442,7 @@ struct test_set : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
 
         auto ne_dst = ne;
@@ -1450,7 +1450,7 @@ struct test_set : public test_case {
             ne_dst[i] *= 2;
         }
         ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, ne_dst.data());
-        ggml_set_param(ctx, dst);
+        ggml_set_param(dst);
         ggml_set_name(dst, "dst");
 
         size_t offset = 0;
@@ -1498,7 +1498,7 @@ struct test_cpy : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
 
         if (_src_use_permute) {
@@ -1536,7 +1536,7 @@ struct test_cont : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
 
         src = ggml_transpose(ctx, src);
@@ -1583,8 +1583,8 @@ struct test_bin_bcast : public test_case {
         // The backward pass supports broadcasting only for GGML_ADD:
         const bool grad_supported = op == ggml_add || ggml_are_same_shape(a, b);
         if (grad_supported) {
-            ggml_set_param(ctx, a);
-            ggml_set_param(ctx, b);
+            ggml_set_param(a);
+            ggml_set_param(b);
         }
 
         ggml_tensor * out = op(ctx, a, b);
@@ -1632,11 +1632,11 @@ struct test_add1 : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * b = ggml_new_tensor_1d(ctx, type, 1);
-        // ggml_set_param(ctx, b); // TODO: implement
+        // ggml_set_param(b); // TODO: implement
         ggml_set_name(b, "b");
 
         ggml_tensor * out = ggml_add1(ctx, a, b);
@@ -1667,7 +1667,7 @@ struct test_scale : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_scale(ctx, a, scale);
@@ -1762,7 +1762,7 @@ struct test_rms_norm : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         if (v) {
@@ -2028,9 +2028,9 @@ struct test_mul_mat : public test_case {
             b = ggml_new_tensor_4d(ctx, type_b, ne_b[per[0]], ne_b[per[1]], ne_b[per[2]], ne_b[per[3]]);
             if (!ggml_is_quantized(type_a)) {
                 if (bs[1] == 1 && nr[1] == 1) {
-                    ggml_set_param(ctx, a);
+                    ggml_set_param(a);
                 }
-                ggml_set_param(ctx, b);
+                ggml_set_param(b);
             }
             ggml_set_name(a, "a");
             ggml_set_name(b, "b");
@@ -2040,22 +2040,29 @@ struct test_mul_mat : public test_case {
             ggml_set_name(a, "a_permuted");
             ggml_set_name(b, "b_permuted");
         } else {
-
             if (v) {
                 a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0],       bs[1]);
                 b = ggml_new_tensor_4d(ctx, type_b, k*2, n, bs[0]*nr[0], bs[1]*nr[1]);
 
+                if (!ggml_is_quantized(type_a)) {
+                    if (bs[1] == 1 && nr[1] == 1) {
+                        ggml_set_param(a);
+                    }
+                    ggml_set_param(b);
+                }
+
                 a = ggml_view_4d(ctx, a, k, m, bs[0],       bs[1],       a->nb[1], a->nb[2], a->nb[3], 0);
                 b = ggml_view_4d(ctx, b, k, n, bs[0]*nr[0], bs[1]*nr[1], b->nb[1], b->nb[2], b->nb[3], 0);
             } else {
                 a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
                 b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
-            }
-            if (!ggml_is_quantized(type_a)) {
-                if (bs[1] == 1 && nr[1] == 1) {
-                    ggml_set_param(ctx, a);
+
+                if (!ggml_is_quantized(type_a)) {
+                    if (bs[1] == 1 && nr[1] == 1) {
+                        ggml_set_param(a);
+                    }
+                    ggml_set_param(b);
                 }
-                ggml_set_param(ctx, b);
             }
             ggml_set_name(a, "a");
             ggml_set_name(b, "b");
@@ -2204,7 +2211,7 @@ struct test_sqr : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_sqr(ctx, a);
@@ -2233,7 +2240,7 @@ struct test_sqrt : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_sqrt(ctx, a);
@@ -2273,7 +2280,7 @@ struct test_log : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_log(ctx, a);
@@ -2309,7 +2316,7 @@ struct test_sin : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_sin(ctx, a);
@@ -2352,7 +2359,7 @@ struct test_cos : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_cos(ctx, a);
@@ -2432,7 +2439,7 @@ struct test_diag_mask_inf : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_diag_mask_inf(ctx, a, n_past);
@@ -2471,7 +2478,7 @@ struct test_soft_max : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * mask = nullptr;
@@ -2553,7 +2560,7 @@ struct test_rope : public test_case {
             auto ne = ne_a; ne[0] *= 2; ne[1] *= 4; ne[2] *= 3;
             a = ggml_new_tensor(ctx, type, 4, ne.data());
             if (forward) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
 
@@ -2562,7 +2569,7 @@ struct test_rope : public test_case {
         } else {
             a = ggml_new_tensor(ctx, type, 4, ne_a.data());
             if (forward) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
         }
@@ -2676,7 +2683,7 @@ struct test_pool2d : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data());
-        ggml_set_param(ctx, input);
+        ggml_set_param(input);
         ggml_set_name(input, "input");
 
         ggml_tensor * out = ggml_pool_2d(ctx, input, pool_type, k0, k1, s0, s1, p0, p1);
@@ -2752,7 +2759,7 @@ struct test_im2col : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data());
-        ggml_set_param(ctx, input);
+        ggml_set_param(input);
         ggml_set_name(input, "input");
 
         ggml_tensor * kernel = ggml_new_tensor(ctx, type_kernel, 4, ne_kernel.data());
@@ -2929,7 +2936,7 @@ struct test_sum : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_sum(ctx, a);
@@ -2958,7 +2965,7 @@ struct test_sum_rows : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_sum_rows(ctx, a);
@@ -2983,7 +2990,7 @@ struct test_mean : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_mean(ctx, a);
@@ -3129,11 +3136,11 @@ struct test_acc : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne_b.data());
-        ggml_set_param(ctx, b);
+        ggml_set_param(b);
         ggml_set_name(b, "b");
 
         ggml_tensor * out = ggml_acc(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], b->nb[1]);
@@ -3370,7 +3377,7 @@ struct test_cross_entropy_loss : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * logits = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, logits);
+        ggml_set_param(logits);
         ggml_set_name(logits, "logits");
 
         ggml_tensor * labels = ggml_new_tensor(ctx, type, 4, ne.data());
@@ -3452,7 +3459,7 @@ struct test_opt_step_adamw : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
-        ggml_set_param(ctx, a); // Despite tensor a having gradients the output tensor will not.
+        ggml_set_param(a); // Despite tensor a having gradients the output tensor will not.
         ggml_set_name(a, "a");
 
         ggml_tensor * grad = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);

tests/test-opt.cpp

@@ -57,7 +57,8 @@ static helper_ctx_data helper_get_ctx_data(
         enum ggml_opt_loss_type loss_type          = GGML_OPT_LOSS_TYPE_SUM) {
     std::vector<ggml_opt_dataset_t> datasets(ndata);
     for (int64_t ndata_shard = 1; ndata_shard <= ndata; ++ndata_shard) {
-        ggml_opt_dataset_t dataset = ggml_opt_dataset_init(ne_datapoint, ne_label, ndata, ndata_shard);
+        ggml_opt_dataset_t dataset = ggml_opt_dataset_init(
+            GGML_TYPE_F32, GGML_TYPE_F32, ne_datapoint, ne_label, ndata, ndata_shard);
 
         float * data   = ggml_get_data_f32(ggml_opt_dataset_data(  dataset));
         float * labels = ggml_get_data_f32(ggml_opt_dataset_labels(dataset));
@@ -74,7 +75,8 @@ static helper_ctx_data helper_get_ctx_data(
         datasets[ndata_shard-1] = dataset;
     }
 
-    ggml_opt_dataset_t dataset_unsupervised = ggml_opt_dataset_init(1, 0, ndata, /*ndata_shard =*/ 1);
+    ggml_opt_dataset_t dataset_unsupervised = ggml_opt_dataset_init(
+        GGML_TYPE_F32, GGML_TYPE_F32, 1, 0, ndata, /*ndata_shard =*/ 1);
 
     float * data = ggml_get_data_f32(ggml_opt_dataset_data(dataset_unsupervised));
 
@@ -113,7 +115,7 @@ static helper_ctx_data helper_get_ctx_data(
 
     struct ggml_tensor * weights = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1);
     ggml_set_name(weights, "weights");
-    ggml_set_param(ctx_static, weights);
+    ggml_set_param(weights);
 
     struct ggml_tensor * intermediary = ggml_add(ctx_compute, inputs, weights);
 
@@ -127,8 +129,11 @@ static helper_ctx_data helper_get_ctx_data(
     GGML_ASSERT(nbatch_logical % nbatch_physical == 0);
     const int32_t opt_period = nbatch_logical / nbatch_physical;
 
-    struct ggml_opt_params opt_params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type);
-    opt_params.opt_period = opt_period;
+    struct ggml_opt_params opt_params = ggml_opt_default_params(backend_sched, loss_type);
+    opt_params.ctx_compute = ctx_compute;
+    opt_params.inputs      = inputs;
+    opt_params.outputs     = outputs;
+    opt_params.opt_period  = opt_period;
     if (!optimizer_defaults) {
         opt_params.get_opt_pars = helper_get_test_opt_pars;
     }
@@ -264,8 +269,9 @@ static std::pair<int, int> test_grad(ggml_backend_sched_t backend_sched, ggml_ba
 
     for (int idata = 0; idata < ndata; ++idata) {
         const float idataf = idata;
+        ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
         ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
-        ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+        ggml_opt_eval(cd.opt_ctx, cd.result);
         ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata, 0, sizeof(float));
     }
 
@@ -334,8 +340,9 @@ static std::pair<int, int> test_forward_backward(
     } else {
         for (int idata = 0; idata < ndata; ++idata) {
             const float idataf = idata;
+            ggml_opt_alloc(cd.opt_ctx, /*backward =*/ false);
             ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
-            ggml_opt_forward(cd.opt_ctx, cd.result);
+            ggml_opt_eval(cd.opt_ctx, cd.result);
             ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
         }
     }
@@ -367,7 +374,8 @@ static std::pair<int, int> test_forward_backward(
     float w0;
     ggml_backend_tensor_get(cd.weights, &w0, 0, sizeof(float));
     for (int i = 0; i < 10; ++i) {
-        ggml_opt_forward_backward(cd.opt_ctx, nullptr);
+        ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
+        ggml_opt_eval(cd.opt_ctx, cd.result);
     }
     ggml_backend_tensor_set(cd.weights, &w0, 0, sizeof(float));
 
@@ -387,8 +395,9 @@ static std::pair<int, int> test_forward_backward(
     } else {
         for (int idata = 0; idata < ndata; ++idata) {
             const float idataf = idata;
+            ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
             ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
-            ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+            ggml_opt_eval(cd.opt_ctx, cd.result);
             ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
         }
     }
@@ -492,14 +501,16 @@ static std::pair<int, int> test_idata_split(ggml_backend_sched_t backend_sched,
             int idata = 0;
             for (; idata < idata_split; ++idata) {
                 const float idataf = idata;
+                ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
                 ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
-                ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+                ggml_opt_eval(cd.opt_ctx, cd.result);
                 ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
             }
             for (; idata < ndata; ++idata) {
                 const float idataf = idata;
+                ggml_opt_alloc(cd.opt_ctx, /*backward =*/ false);
                 ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
-                ggml_opt_forward(cd.opt_ctx, cd.result2);
+                ggml_opt_eval(cd.opt_ctx, cd.result2);
                 ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
             }
         }
@@ -573,7 +584,6 @@ static std::pair<int, int> test_gradient_accumulation(
 
     struct helper_ctx_data cd = helper_get_ctx_data(
         backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false, /*nbatch_logical =*/ 6, nbatch_physical, loss_type);
-    struct ggml_tensor * loss = ggml_opt_loss(cd.opt_ctx);
 
     std::vector<float> grad_history(ndata);
     for (int64_t idata = 0; idata < ndata; ++idata) {
@@ -584,15 +594,17 @@ static std::pair<int, int> test_gradient_accumulation(
         if (nbatch_physical == 1) {
             for (int idata = 0; idata < ndata; ++idata) {
                 const float idataf = idata;
+                ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
                 ggml_backend_tensor_set(cd.inputs, &idataf, 0, 1*sizeof(float));
-                ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+                ggml_opt_eval(cd.opt_ctx, cd.result);
                 ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata, 0, 1*sizeof(float));
             }
         } else if (nbatch_physical == 2) {
             for (int idata = 0; idata < ndata; idata += 2) {
                 const float idataf[2] = {float(idata + 0), float(idata + 1)};
+                ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
                 ggml_backend_tensor_set(cd.inputs, idataf, 0, 2*sizeof(float));
-                ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+                ggml_opt_eval(cd.opt_ctx, cd.result);
 
                 grad_history[idata + 0] = 0.0f;
                 ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata + 1, 0, 1*sizeof(float));
@@ -617,7 +629,7 @@ static std::pair<int, int> test_gradient_accumulation(
                 }
                 subtest_ok = subtest_ok && almost_equal(grad_history[1], 2.0, atol);
                 subtest_ok = subtest_ok && almost_equal(grad_history[3], 4.0, atol);
-                subtest_ok = subtest_ok && almost_equal(grad_history[5], 0.0, atol);
+                subtest_ok = subtest_ok && almost_equal(grad_history[5], 6.0, atol);
             } else if (loss_type == GGML_OPT_LOSS_TYPE_MEAN) {
                 if (nbatch_physical == 1) {
                     subtest_ok = subtest_ok && almost_equal(grad_history[0], 1.0/ndata, atol);
@@ -630,7 +642,7 @@ static std::pair<int, int> test_gradient_accumulation(
                 }
                 subtest_ok = subtest_ok && almost_equal(grad_history[1], 2.0/ndata, atol);
                 subtest_ok = subtest_ok && almost_equal(grad_history[3], 4.0/ndata, atol);
-                subtest_ok = subtest_ok && almost_equal(grad_history[5], 0.0/ndata, atol);
+                subtest_ok = subtest_ok && almost_equal(grad_history[5], 6.0/ndata, atol);
             } else {
                 GGML_ASSERT(false);
             }
@@ -692,7 +704,8 @@ static std::pair<int, int> test_regression(ggml_backend_sched_t backend_sched, g
     std::mt19937 gen(12345);
     std::normal_distribution<float> nd{0.0f, 0.1f};
 
-    ggml_opt_dataset_t dataset = ggml_opt_dataset_init(1, 1, ndata_regression, ndata_regression);
+    ggml_opt_dataset_t dataset = ggml_opt_dataset_init(
+        GGML_TYPE_F32, GGML_TYPE_F32, 1, 1, ndata_regression, ndata_regression);
 
     float * data   = ggml_get_data_f32(ggml_opt_dataset_data(  dataset));
     float * labels = ggml_get_data_f32(ggml_opt_dataset_labels(dataset));
@@ -733,15 +746,14 @@ static std::pair<int, int> test_regression(ggml_backend_sched_t backend_sched, g
 
     struct ggml_tensor * a = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1);
     ggml_set_name(a, "a");
-    ggml_set_param(ctx_static, a);
+    ggml_set_param(a);
 
     struct ggml_tensor * b = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1);
     ggml_set_name(b, "b");
-    ggml_set_param(ctx_static, b);
+    ggml_set_param(b);
 
     struct ggml_tensor * f = ggml_add(ctx_compute, ggml_mul(ctx_compute, x, a), b);
     ggml_set_name(f, "f");
-    ggml_set_param(ctx_static, f);
 
     ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx_static, backend);
     const float a0 = 1.0f;

tools/llama-bench/README.md

@@ -20,10 +20,20 @@ Performance testing tool for llama.cpp.
 ## Syntax
 
 ```
-usage: ./llama-bench [options]
+usage: llama-bench [options]
 
 options:
   -h, --help
+  --numa <distribute|isolate|numactl>       numa mode (default: disabled)
+  -r, --repetitions <n>                     number of times to repeat each test (default: 5)
+  --prio <0|1|2|3>                          process/thread priority (default: 0)
+  --delay <0...N> (seconds)                 delay between each test (default: 0)
+  -o, --output <csv|json|jsonl|md|sql>      output format printed to stdout (default: md)
+  -oe, --output-err <csv|json|jsonl|md|sql> output format printed to stderr (default: none)
+  -v, --verbose                             verbose output
+  --progress                                print test progress indicators
+
+test parameters:
   -m, --model <filename>                    (default: models/7B/ggml-model-q4_0.gguf)
   -p, --n-prompt <n>                        (default: 512)
   -n, --n-gen <n>                           (default: 128)
@@ -33,28 +43,27 @@ options:
   -ub, --ubatch-size <n>                    (default: 512)
   -ctk, --cache-type-k <t>                  (default: f16)
   -ctv, --cache-type-v <t>                  (default: f16)
-  -t, --threads <n>                         (default: 8)
+  -dt, --defrag-thold <f>                   (default: -1)
+  -t, --threads <n>                         (default: system dependent)
   -C, --cpu-mask <hex,hex>                  (default: 0x0)
   --cpu-strict <0|1>                        (default: 0)
   --poll <0...100>                          (default: 50)
   -ngl, --n-gpu-layers <n>                  (default: 99)
-  -rpc, --rpc <rpc_servers>                 (default: )
+  -rpc, --rpc <rpc_servers>                 (default: none)
   -sm, --split-mode <none|layer|row>        (default: layer)
   -mg, --main-gpu <i>                       (default: 0)
   -nkvo, --no-kv-offload <0|1>              (default: 0)
   -fa, --flash-attn <0|1>                   (default: 0)
   -mmp, --mmap <0|1>                        (default: 1)
-  --numa <distribute|isolate|numactl>       (default: disabled)
   -embd, --embeddings <0|1>                 (default: 0)
   -ts, --tensor-split <ts0/ts1/..>          (default: 0)
-  -r, --repetitions <n>                     (default: 5)
-  --prio <0|1|2|3>                          (default: 0)
-  --delay <0...N> (seconds)                 (default: 0)
-  -o, --output <csv|json|jsonl|md|sql>      (default: md)
-  -oe, --output-err <csv|json|jsonl|md|sql> (default: none)
-  -v, --verbose                             (default: 0)
+  -ot --override-tensors <tensor name pattern>=<buffer type>;...
+                                            (default: disabled)
+  -nopo, --no-op-offload <0|1>              (default: 0)
 
-Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
+Multiple values can be given for each parameter by separating them with ','
+or by specifying the parameter multiple times. Ranges can be given as
+'first-last' or 'first-last+step' or 'first-last*mult'.
 ```
 
 llama-bench can perform three types of tests:

tools/llama-bench/llama-bench.cpp

@@ -195,6 +195,46 @@ static std::string pair_str(const std::pair<int, int> & p) {
     return buf;
 }
 
+static std::vector<int> parse_int_range(const std::string & s) {
+    // first[-last[(+|*)step]]
+    std::regex range_regex(R"(^(\d+)(?:-(\d+)(?:([\+|\*])(\d+))?)?(?:,|$))");
+
+    std::smatch match;
+    std::string::const_iterator search_start(s.cbegin());
+    std::vector<int> result;
+    while (std::regex_search(search_start, s.cend(), match, range_regex)) {
+        int  first = std::stoi(match[1]);
+        int  last  = match[2].matched ? std::stoi(match[2]) : first;
+        char op    = match[3].matched ? match[3].str()[0] : '+';
+        int  step  = match[4].matched ? std::stoi(match[4]) : 1;
+
+        for (int i = first; i <= last;) {
+            result.push_back(i);
+
+            int prev_i = i;
+
+            if (op == '+') {
+                i += step;
+            } else if (op == '*') {
+                i *= step;
+            } else {
+                throw std::invalid_argument("invalid range format");
+            }
+
+            if (i <= prev_i) {
+                throw std::invalid_argument("invalid range");
+            }
+        }
+        search_start = match.suffix().first;
+    }
+
+    if (search_start != s.cend()) {
+        throw std::invalid_argument("invalid range format");
+    }
+
+    return result;
+}
+
 struct cmd_params {
     std::vector<std::string>         model;
     std::vector<int>                 n_prompt;
@@ -205,6 +245,7 @@ struct cmd_params {
     std::vector<int>                 n_ubatch;
     std::vector<ggml_type>           type_k;
     std::vector<ggml_type>           type_v;
+    std::vector<float>               defrag_thold;
     std::vector<int>                 n_threads;
     std::vector<std::string>         cpu_mask;
     std::vector<bool>                cpu_strict;
@@ -240,6 +281,7 @@ static const cmd_params cmd_params_defaults = {
     /* n_ubatch             */ { 512 },
     /* type_k               */ { GGML_TYPE_F16 },
     /* type_v               */ { GGML_TYPE_F16 },
+    /* defrag_thold         */ { -1.0f },
     /* n_threads            */ { cpu_get_num_math() },
     /* cpu_mask             */ { "0x0" },
     /* cpu_strict           */ { false },
@@ -251,7 +293,7 @@ static const cmd_params cmd_params_defaults = {
     /* no_kv_offload        */ { false },
     /* flash_attn           */ { false },
     /* tensor_split         */ { std::vector<float>(llama_max_devices(), 0.0f) },
-    /* tensor_buft_overrides*/ { std::vector<llama_model_tensor_buft_override>{{nullptr,nullptr}} },
+    /* tensor_buft_overrides*/ { std::vector<llama_model_tensor_buft_override>{ { nullptr, nullptr } } },
     /* use_mmap             */ { true },
     /* embeddings           */ { false },
     /* no_op_offload        */ { false },
@@ -270,13 +312,29 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("\n");
     printf("options:\n");
     printf("  -h, --help\n");
+    printf("  --numa <distribute|isolate|numactl>       numa mode (default: disabled)\n");
+    printf("  -r, --repetitions <n>                     number of times to repeat each test (default: %d)\n",
+           cmd_params_defaults.reps);
+    printf("  --prio <0|1|2|3>                          process/thread priority (default: %d)\n",
+           cmd_params_defaults.prio);
+    printf("  --delay <0...N> (seconds)                 delay between each test (default: %d)\n",
+           cmd_params_defaults.delay);
+    printf("  -o, --output <csv|json|jsonl|md|sql>      output format printed to stdout (default: %s)\n",
+           output_format_str(cmd_params_defaults.output_format));
+    printf("  -oe, --output-err <csv|json|jsonl|md|sql> output format printed to stderr (default: %s)\n",
+           output_format_str(cmd_params_defaults.output_format_stderr));
+    printf("  -v, --verbose                             verbose output\n");
+    printf("  --progress                                print test progress indicators\n");
+    printf("\n");
+    printf("test parameters:\n");
     printf("  -m, --model <filename>                    (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
     printf("  -p, --n-prompt <n>                        (default: %s)\n",
            join(cmd_params_defaults.n_prompt, ",").c_str());
     printf("  -n, --n-gen <n>                           (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
     printf("  -pg <pp,tg>                               (default: %s)\n",
            join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
-    printf("  -d, --n-depth <n>                         (default: %s)\n", join(cmd_params_defaults.n_depth, ",").c_str());
+    printf("  -d, --n-depth <n>                         (default: %s)\n",
+           join(cmd_params_defaults.n_depth, ",").c_str());
     printf("  -b, --batch-size <n>                      (default: %s)\n",
            join(cmd_params_defaults.n_batch, ",").c_str());
     printf("  -ub, --ubatch-size <n>                    (default: %s)\n",
@@ -285,6 +343,8 @@ static void print_usage(int /* argc */, char ** argv) {
            join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
     printf("  -ctv, --cache-type-v <t>                  (default: %s)\n",
            join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
+    printf("  -dt, --defrag-thold <f>                   (default: %s)\n",
+           join(cmd_params_defaults.defrag_thold, ",").c_str());
     printf("  -t, --threads <n>                         (default: %s)\n",
            join(cmd_params_defaults.n_threads, ",").c_str());
     printf("  -C, --cpu-mask <hex,hex>                  (default: %s)\n",
@@ -308,25 +368,17 @@ static void print_usage(int /* argc */, char ** argv) {
            join(cmd_params_defaults.flash_attn, ",").c_str());
     printf("  -mmp, --mmap <0|1>                        (default: %s)\n",
            join(cmd_params_defaults.use_mmap, ",").c_str());
-    printf("  --numa <distribute|isolate|numactl>       (default: disabled)\n");
     printf("  -embd, --embeddings <0|1>                 (default: %s)\n",
            join(cmd_params_defaults.embeddings, ",").c_str());
     printf("  -ts, --tensor-split <ts0/ts1/..>          (default: 0)\n");
-    printf("  -ot --override-tensors <tensor name pattern>=<buffer type>;... (default: disabled)\n");
-    printf("  -nopo, --no-op-offload <i>                (default: 0)\n");
-    printf("  -r, --repetitions <n>                     (default: %d)\n", cmd_params_defaults.reps);
-    printf("  --prio <0|1|2|3>                          (default: %d)\n", cmd_params_defaults.prio);
-    printf("  --delay <0...N> (seconds)                 (default: %d)\n", cmd_params_defaults.delay);
-    printf("  -o, --output <csv|json|jsonl|md|sql>      (default: %s)\n",
-           output_format_str(cmd_params_defaults.output_format));
-    printf("  -oe, --output-err <csv|json|jsonl|md|sql> (default: %s)\n",
-           output_format_str(cmd_params_defaults.output_format_stderr));
-    printf("  -v, --verbose                             (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
-    printf("  --progress                                (default: %s)\n", cmd_params_defaults.progress ? "1" : "0");
+    printf("  -ot --override-tensors <tensor name pattern>=<buffer type>;...\n");
+    printf("                                            (default: disabled)\n");
+    printf("  -nopo, --no-op-offload <0|1>              (default: 0)\n");
     printf("\n");
     printf(
-        "Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter "
-        "multiple times.\n");
+        "Multiple values can be given for each parameter by separating them with ','\n"
+        "or by specifying the parameter multiple times. Ranges can be given as\n"
+        "'first-last' or 'first-last+step' or 'first-last*mult'.\n");
 }
 
 static ggml_type ggml_type_from_name(const std::string & s) {
@@ -380,186 +432,197 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             std::replace(arg.begin(), arg.end(), '_', '-');
         }
 
-        if (arg == "-h" || arg == "--help") {
-            print_usage(argc, argv);
-            exit(0);
-        } else if (arg == "-m" || arg == "--model") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<std::string>(argv[i], split_delim);
-            params.model.insert(params.model.end(), p.begin(), p.end());
-        } else if (arg == "-p" || arg == "--n-prompt") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<int>(argv[i], split_delim);
-            params.n_prompt.insert(params.n_prompt.end(), p.begin(), p.end());
-        } else if (arg == "-n" || arg == "--n-gen") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<int>(argv[i], split_delim);
-            params.n_gen.insert(params.n_gen.end(), p.begin(), p.end());
-        } else if (arg == "-pg") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<std::string>(argv[i], ',');
-            if (p.size() != 2) {
-                invalid_param = true;
-                break;
-            }
-            params.n_pg.push_back({ std::stoi(p[0]), std::stoi(p[1]) });
-        } else if (arg == "-d" || arg == "--n-depth") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<int>(argv[i], split_delim);
-            params.n_depth.insert(params.n_depth.end(), p.begin(), p.end());
-        } else if (arg == "-b" || arg == "--batch-size") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<int>(argv[i], split_delim);
-            params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
-        } else if (arg == "-ub" || arg == "--ubatch-size") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<int>(argv[i], split_delim);
-            params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end());
-        } else if (arg == "-ctk" || arg == "--cache-type-k") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto                   p = string_split<std::string>(argv[i], split_delim);
-            std::vector<ggml_type> types;
-            for (const auto & t : p) {
-                ggml_type gt = ggml_type_from_name(t);
-                if (gt == GGML_TYPE_COUNT) {
+        try {
+            if (arg == "-h" || arg == "--help") {
+                print_usage(argc, argv);
+                exit(0);
+            } else if (arg == "-m" || arg == "--model") {
+                if (++i >= argc) {
                     invalid_param = true;
                     break;
                 }
-                types.push_back(gt);
-            }
-            if (invalid_param) {
-                break;
-            }
-            params.type_k.insert(params.type_k.end(), types.begin(), types.end());
-        } else if (arg == "-ctv" || arg == "--cache-type-v") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto                   p = string_split<std::string>(argv[i], split_delim);
-            std::vector<ggml_type> types;
-            for (const auto & t : p) {
-                ggml_type gt = ggml_type_from_name(t);
-                if (gt == GGML_TYPE_COUNT) {
+                auto p = string_split<std::string>(argv[i], split_delim);
+                params.model.insert(params.model.end(), p.begin(), p.end());
+            } else if (arg == "-p" || arg == "--n-prompt") {
+                if (++i >= argc) {
                     invalid_param = true;
                     break;
                 }
-                types.push_back(gt);
-            }
-            if (invalid_param) {
-                break;
-            }
-            params.type_v.insert(params.type_v.end(), types.begin(), types.end());
-        } else if (arg == "-t" || arg == "--threads") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<int>(argv[i], split_delim);
-            params.n_threads.insert(params.n_threads.end(), p.begin(), p.end());
-        } else if (arg == "-C" || arg == "--cpu-mask") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<std::string>(argv[i], split_delim);
-            params.cpu_mask.insert(params.cpu_mask.end(), p.begin(), p.end());
-        } else if (arg == "--cpu-strict") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<bool>(argv[i], split_delim);
-            params.cpu_strict.insert(params.cpu_strict.end(), p.begin(), p.end());
-        } else if (arg == "--poll") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<int>(argv[i], split_delim);
-            params.poll.insert(params.poll.end(), p.begin(), p.end());
-        } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<int>(argv[i], split_delim);
-            params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
-        } else if (llama_supports_rpc() && (arg == "-rpc" || arg == "--rpc")) {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.rpc_servers.push_back(argv[i]);
-        } else if (arg == "-sm" || arg == "--split-mode") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto                          p = string_split<std::string>(argv[i], split_delim);
-            std::vector<llama_split_mode> modes;
-            for (const auto & m : p) {
-                llama_split_mode mode;
-                if (m == "none") {
-                    mode = LLAMA_SPLIT_MODE_NONE;
-                } else if (m == "layer") {
-                    mode = LLAMA_SPLIT_MODE_LAYER;
-                } else if (m == "row") {
-                    mode = LLAMA_SPLIT_MODE_ROW;
-                } else {
+                auto p = parse_int_range(argv[i]);
+                params.n_prompt.insert(params.n_prompt.end(), p.begin(), p.end());
+            } else if (arg == "-n" || arg == "--n-gen") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = parse_int_range(argv[i]);
+                params.n_gen.insert(params.n_gen.end(), p.begin(), p.end());
+            } else if (arg == "-pg") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<std::string>(argv[i], ',');
+                if (p.size() != 2) {
+                    invalid_param = true;
+                    break;
+                }
+                params.n_pg.push_back({ std::stoi(p[0]), std::stoi(p[1]) });
+            } else if (arg == "-d" || arg == "--n-depth") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = parse_int_range(argv[i]);
+                params.n_depth.insert(params.n_depth.end(), p.begin(), p.end());
+            } else if (arg == "-b" || arg == "--batch-size") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = parse_int_range(argv[i]);
+                params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
+            } else if (arg == "-ub" || arg == "--ubatch-size") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = parse_int_range(argv[i]);
+                params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end());
+            } else if (arg == "-ctk" || arg == "--cache-type-k") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<std::string>(argv[i], split_delim);
+
+                std::vector<ggml_type> types;
+                for (const auto & t : p) {
+                    ggml_type gt = ggml_type_from_name(t);
+                    if (gt == GGML_TYPE_COUNT) {
+                        invalid_param = true;
+                        break;
+                    }
+                    types.push_back(gt);
+                }
+                if (invalid_param) {
+                    break;
+                }
+                params.type_k.insert(params.type_k.end(), types.begin(), types.end());
+            } else if (arg == "-ctv" || arg == "--cache-type-v") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<std::string>(argv[i], split_delim);
+
+                std::vector<ggml_type> types;
+                for (const auto & t : p) {
+                    ggml_type gt = ggml_type_from_name(t);
+                    if (gt == GGML_TYPE_COUNT) {
+                        invalid_param = true;
+                        break;
+                    }
+                    types.push_back(gt);
+                }
+                if (invalid_param) {
+                    break;
+                }
+                params.type_v.insert(params.type_v.end(), types.begin(), types.end());
+            } else if (arg == "-dt" || arg == "--defrag-thold") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<float>(argv[i], split_delim);
+                params.defrag_thold.insert(params.defrag_thold.end(), p.begin(), p.end());
+            } else if (arg == "-t" || arg == "--threads") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = parse_int_range(argv[i]);
+                params.n_threads.insert(params.n_threads.end(), p.begin(), p.end());
+            } else if (arg == "-C" || arg == "--cpu-mask") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<std::string>(argv[i], split_delim);
+                params.cpu_mask.insert(params.cpu_mask.end(), p.begin(), p.end());
+            } else if (arg == "--cpu-strict") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<bool>(argv[i], split_delim);
+                params.cpu_strict.insert(params.cpu_strict.end(), p.begin(), p.end());
+            } else if (arg == "--poll") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = parse_int_range(argv[i]);
+                params.poll.insert(params.poll.end(), p.begin(), p.end());
+            } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = parse_int_range(argv[i]);
+                params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
+            } else if (llama_supports_rpc() && (arg == "-rpc" || arg == "--rpc")) {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                params.rpc_servers.push_back(argv[i]);
+            } else if (arg == "-sm" || arg == "--split-mode") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<std::string>(argv[i], split_delim);
+
+                std::vector<llama_split_mode> modes;
+                for (const auto & m : p) {
+                    llama_split_mode mode;
+                    if (m == "none") {
+                        mode = LLAMA_SPLIT_MODE_NONE;
+                    } else if (m == "layer") {
+                        mode = LLAMA_SPLIT_MODE_LAYER;
+                    } else if (m == "row") {
+                        mode = LLAMA_SPLIT_MODE_ROW;
+                    } else {
+                        invalid_param = true;
+                        break;
+                    }
+                    modes.push_back(mode);
+                }
+                if (invalid_param) {
+                    break;
+                }
+                params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
+            } else if (arg == "-mg" || arg == "--main-gpu") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                params.main_gpu = parse_int_range(argv[i]);
+            } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<bool>(argv[i], split_delim);
+                params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
+            } else if (arg == "--numa") {
+                if (++i >= argc) {
                     invalid_param = true;
                     break;
                 }
-                modes.push_back(mode);
-            }
-            if (invalid_param) {
-                break;
-            }
-            params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
-        } else if (arg == "-mg" || arg == "--main-gpu") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.main_gpu = string_split<int>(argv[i], split_delim);
-        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<bool>(argv[i], split_delim);
-            params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
-        } else if (arg == "--numa") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            } else {
                 std::string value(argv[i]);
-                /**/ if (value == "distribute" || value == "") {
+                if (value == "distribute" || value == "") {
                     params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE;
                 } else if (value == "isolate") {
                     params.numa = GGML_NUMA_STRATEGY_ISOLATE;
@@ -569,177 +632,182 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                     invalid_param = true;
                     break;
                 }
-            }
-        } else if (arg == "-fa" || arg == "--flash-attn") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<bool>(argv[i], split_delim);
-            params.flash_attn.insert(params.flash_attn.end(), p.begin(), p.end());
-        } else if (arg == "-mmp" || arg == "--mmap") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<bool>(argv[i], split_delim);
-            params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end());
-        } else if (arg == "-embd" || arg == "--embeddings") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<bool>(argv[i], split_delim);
-            params.embeddings.insert(params.embeddings.end(), p.begin(), p.end());
-        } else if (arg == "-nopo" || arg == "--no-op-offload") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<bool>(argv[i], split_delim);
-            params.no_op_offload.insert(params.no_op_offload.end(), p.begin(), p.end());
-        } else if (arg == "-ts" || arg == "--tensor-split") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            for (auto ts : string_split<std::string>(argv[i], split_delim)) {
-                // split string by ; and /
-                const std::regex           regex{ R"([;/]+)" };
-                std::sregex_token_iterator it{ ts.begin(), ts.end(), regex, -1 };
-                std::vector<std::string>   split_arg{ it, {} };
-                GGML_ASSERT(split_arg.size() <= llama_max_devices());
+            } else if (arg == "-fa" || arg == "--flash-attn") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<bool>(argv[i], split_delim);
+                params.flash_attn.insert(params.flash_attn.end(), p.begin(), p.end());
+            } else if (arg == "-mmp" || arg == "--mmap") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<bool>(argv[i], split_delim);
+                params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end());
+            } else if (arg == "-embd" || arg == "--embeddings") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<bool>(argv[i], split_delim);
+                params.embeddings.insert(params.embeddings.end(), p.begin(), p.end());
+            } else if (arg == "-nopo" || arg == "--no-op-offload") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto p = string_split<bool>(argv[i], split_delim);
+                params.no_op_offload.insert(params.no_op_offload.end(), p.begin(), p.end());
+            } else if (arg == "-ts" || arg == "--tensor-split") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                for (auto ts : string_split<std::string>(argv[i], split_delim)) {
+                    // split string by ; and /
+                    const std::regex           regex{ R"([;/]+)" };
+                    std::sregex_token_iterator it{ ts.begin(), ts.end(), regex, -1 };
+                    std::vector<std::string>   split_arg{ it, {} };
+                    GGML_ASSERT(split_arg.size() <= llama_max_devices());
 
-                std::vector<float> tensor_split(llama_max_devices());
-                for (size_t i = 0; i < llama_max_devices(); ++i) {
-                    if (i < split_arg.size()) {
-                        tensor_split[i] = std::stof(split_arg[i]);
-                    } else {
-                        tensor_split[i] = 0.0f;
+                    std::vector<float> tensor_split(llama_max_devices());
+                    for (size_t i = 0; i < llama_max_devices(); ++i) {
+                        if (i < split_arg.size()) {
+                            tensor_split[i] = std::stof(split_arg[i]);
+                        } else {
+                            tensor_split[i] = 0.0f;
+                        }
+                    }
+                    params.tensor_split.push_back(tensor_split);
+                }
+            } else if (arg == "-ot" || arg == "--override-tensor") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                auto value = argv[i];
+                /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
+                if (buft_list.empty()) {
+                    // enumerate all the devices and add their buffer types to the list
+                    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+                        auto * dev = ggml_backend_dev_get(i);
+                        auto * buft = ggml_backend_dev_buffer_type(dev);
+                        if (buft) {
+                            buft_list[ggml_backend_buft_name(buft)] = buft;
+                        }
                     }
                 }
-                params.tensor_split.push_back(tensor_split);
-            }
-        } else if (arg == "-ot" || arg == "--override-tensor") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto value = argv[i];
-            /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
-            if (buft_list.empty()) {
-                // enumerate all the devices and add their buffer types to the list
-                for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-                    auto * dev = ggml_backend_dev_get(i);
-                    auto * buft = ggml_backend_dev_buffer_type(dev);
-                    if (buft) {
-                        buft_list[ggml_backend_buft_name(buft)] = buft;
+                auto override_group_span_len = std::strcspn(value, ",");
+                bool last_group = false;
+                do {
+                    if (override_group_span_len == 0) {
+                        // Adds an empty override-tensors for an empty span
+                        params.tensor_buft_overrides.push_back({{}});
+                        if (value[override_group_span_len] == '\0') {
+                            value = &value[override_group_span_len];
+                            last_group = true;
+                        } else {
+                            value = &value[override_group_span_len + 1];
+                            override_group_span_len = std::strcspn(value, ",");
+                        }
+                        continue;
                     }
-                }
-            }
-            auto override_group_span_len = std::strcspn(value, ",");
-            bool last_group = false;
-            do {
-                if (override_group_span_len == 0) {
-                    // Adds an empty override-tensors for an empty span
-                    params.tensor_buft_overrides.push_back({{}});
+                    // Stamps null terminators into the argv
+                    // value for this option to avoid the
+                    // memory leak present in the implementation
+                    // over in arg.cpp. Acceptable because we
+                    // only parse these args once in this program.
+                    auto override_group = value;
                     if (value[override_group_span_len] == '\0') {
                         value = &value[override_group_span_len];
                         last_group = true;
                     } else {
+                        value[override_group_span_len] = '\0';
                         value = &value[override_group_span_len + 1];
-                        override_group_span_len = std::strcspn(value, ",");
                     }
-                    continue;
-                }
-                // Stamps null terminators into the argv
-                // value for this option to avoid the
-                // memory leak present in the implementation
-                // over in arg.cpp. Acceptable because we
-                // only parse these args once in this program.
-                auto override_group = value;
-                if (value[override_group_span_len] == '\0') {
-                    value = &value[override_group_span_len];
-                    last_group = true;
-                } else {
-                    value[override_group_span_len] = '\0';
-                    value = &value[override_group_span_len + 1];
-                }
-                std::vector<llama_model_tensor_buft_override> group_tensor_buft_overrides{};
-                auto override_span_len = std::strcspn(override_group, ";");
-                while (override_span_len > 0) {
-                    auto override = override_group;
-                    if (override_group[override_span_len] != '\0') {
-                        override_group[override_span_len] = '\0';
-                        override_group = &override_group[override_span_len + 1];
-                    } else {
-                        override_group = &override_group[override_span_len];
-                    }
-                    auto tensor_name_span_len = std::strcspn(override, "=");
-                    if (tensor_name_span_len >= override_span_len) {
-                        invalid_param = true;
-                        break;
-                    }
-                    override[tensor_name_span_len] = '\0';
-                    auto tensor_name = override;
-                    auto buffer_type = &override[tensor_name_span_len + 1];
-                    if (buft_list.find(buffer_type) == buft_list.end()) {
-                        printf("Available buffer types:\n");
-                        for (const auto & it : buft_list) {
-                            printf("  %s\n", ggml_backend_buft_name(it.second));
+                    std::vector<llama_model_tensor_buft_override> group_tensor_buft_overrides{};
+                    auto override_span_len = std::strcspn(override_group, ";");
+                    while (override_span_len > 0) {
+                        auto override = override_group;
+                        if (override_group[override_span_len] != '\0') {
+                            override_group[override_span_len] = '\0';
+                            override_group = &override_group[override_span_len + 1];
+                        } else {
+                            override_group = &override_group[override_span_len];
                         }
-                        invalid_param = true;
+                        auto tensor_name_span_len = std::strcspn(override, "=");
+                        if (tensor_name_span_len >= override_span_len) {
+                            invalid_param = true;
+                            break;
+                        }
+                        override[tensor_name_span_len] = '\0';
+                        auto tensor_name = override;
+                        auto buffer_type = &override[tensor_name_span_len + 1];
+                        if (buft_list.find(buffer_type) == buft_list.end()) {
+                            printf("Available buffer types:\n");
+                            for (const auto & it : buft_list) {
+                                printf("  %s\n", ggml_backend_buft_name(it.second));
+                            }
+                            invalid_param = true;
+                            break;
+                        }
+                        group_tensor_buft_overrides.push_back({tensor_name, buft_list.at(buffer_type)});
+                        override_span_len = std::strcspn(override_group, ";");
+                    }
+                    if (invalid_param) {
                         break;
                     }
-                    group_tensor_buft_overrides.push_back({tensor_name, buft_list.at(buffer_type)});
-                    override_span_len = std::strcspn(override_group, ";");
-                }
-                if (invalid_param) {
+                    group_tensor_buft_overrides.push_back({nullptr,nullptr});
+                    params.tensor_buft_overrides.push_back(group_tensor_buft_overrides);
+                    override_group_span_len = std::strcspn(value, ",");
+                } while (!last_group);
+            } else if (arg == "-r" || arg == "--repetitions") {
+                if (++i >= argc) {
+                    invalid_param = true;
                     break;
                 }
-                group_tensor_buft_overrides.push_back({nullptr,nullptr});
-                params.tensor_buft_overrides.push_back(group_tensor_buft_overrides);
-                override_group_span_len = std::strcspn(value, ",");
-            } while (!last_group);
-        } else if (arg == "-r" || arg == "--repetitions") {
-            if (++i >= argc) {
+                params.reps = std::stoi(argv[i]);
+            } else if (arg == "--prio") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                params.prio = (enum ggml_sched_priority) std::stoi(argv[i]);
+            } else if (arg == "--delay") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                params.delay = std::stoi(argv[i]);
+            } else if (arg == "-o" || arg == "--output") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                invalid_param = !output_format_from_str(argv[i], params.output_format);
+            } else if (arg == "-oe" || arg == "--output-err") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                invalid_param = !output_format_from_str(argv[i], params.output_format_stderr);
+            } else if (arg == "-v" || arg == "--verbose") {
+                params.verbose = true;
+            } else if (arg == "--progress") {
+                params.progress = true;
+            } else {
                 invalid_param = true;
                 break;
             }
-            params.reps = std::stoi(argv[i]);
-        } else if (arg == "--prio") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.prio = (enum ggml_sched_priority) std::stoi(argv[i]);
-        } else if (arg == "--delay") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.delay = std::stoi(argv[i]);
-        } else if (arg == "-o" || arg == "--output") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            invalid_param = !output_format_from_str(argv[i], params.output_format);
-        } else if (arg == "-oe" || arg == "--output-err") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            invalid_param = !output_format_from_str(argv[i], params.output_format_stderr);
-        } else if (arg == "-v" || arg == "--verbose") {
-            params.verbose = true;
-        } else if (arg == "--progress") {
-            params.progress = true;
-        } else {
+        } catch (const std::exception & e) {
+            fprintf(stderr, "error: %s\n", e.what());
             invalid_param = true;
             break;
         }
     }
+
     if (invalid_param) {
         fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
         print_usage(argc, argv);
@@ -774,6 +842,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.type_v.empty()) {
         params.type_v = cmd_params_defaults.type_v;
     }
+    if (params.defrag_thold.empty()) {
+        params.defrag_thold = cmd_params_defaults.defrag_thold;
+    }
     if (params.n_gpu_layers.empty()) {
         params.n_gpu_layers = cmd_params_defaults.n_gpu_layers;
     }
@@ -832,6 +903,7 @@ struct cmd_params_instance {
     int                n_ubatch;
     ggml_type          type_k;
     ggml_type          type_v;
+    float              defrag_thold;
     int                n_threads;
     std::string        cpu_mask;
     bool               cpu_strict;
@@ -908,15 +980,16 @@ struct cmd_params_instance {
     llama_context_params to_llama_cparams() const {
         llama_context_params cparams = llama_context_default_params();
 
-        cparams.n_ctx       = n_prompt + n_gen + n_depth;
-        cparams.n_batch     = n_batch;
-        cparams.n_ubatch    = n_ubatch;
-        cparams.type_k      = type_k;
-        cparams.type_v      = type_v;
-        cparams.offload_kqv = !no_kv_offload;
-        cparams.flash_attn  = flash_attn;
-        cparams.embeddings  = embeddings;
-        cparams.op_offload  = !no_op_offload;
+        cparams.n_ctx        = n_prompt + n_gen + n_depth;
+        cparams.n_batch      = n_batch;
+        cparams.n_ubatch     = n_ubatch;
+        cparams.type_k       = type_k;
+        cparams.type_v       = type_v;
+        cparams.defrag_thold = defrag_thold;
+        cparams.offload_kqv  = !no_kv_offload;
+        cparams.flash_attn   = flash_attn;
+        cparams.embeddings   = embeddings;
+        cparams.op_offload   = !no_op_offload;
 
         return cparams;
     }
@@ -941,6 +1014,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & nub : params.n_ubatch)
     for (const auto & tk : params.type_k)
     for (const auto & tv : params.type_v)
+    for (const auto & defrag_thold : params.defrag_thold)
     for (const auto & nkvo : params.no_kv_offload)
     for (const auto & fa : params.flash_attn)
     for (const auto & nt : params.n_threads)
@@ -961,6 +1035,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
+                /* .defrag_thold = */ defrag_thold,
                 /* .n_threads    = */ nt,
                 /* .cpu_mask     = */ cm,
                 /* .cpu_strict   = */ cs,
@@ -993,6 +1068,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
+                /* .defrag_thold = */ defrag_thold,
                 /* .n_threads    = */ nt,
                 /* .cpu_mask     = */ cm,
                 /* .cpu_strict   = */ cs,
@@ -1025,6 +1101,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
+                /* .defrag_thold = */ defrag_thold,
                 /* .n_threads    = */ nt,
                 /* .cpu_mask     = */ cm,
                 /* .cpu_strict   = */ cs,
@@ -1066,6 +1143,7 @@ struct test {
     int                      poll;
     ggml_type                type_k;
     ggml_type                type_v;
+    float                    defrag_thold;
     int                      n_gpu_layers;
     llama_split_mode         split_mode;
     int                      main_gpu;
@@ -1100,6 +1178,7 @@ struct test {
         poll           = inst.poll;
         type_k         = inst.type_k;
         type_v         = inst.type_v;
+        defrag_thold   = inst.defrag_thold;
         n_gpu_layers   = inst.n_gpu_layers;
         split_mode     = inst.split_mode;
         main_gpu       = inst.main_gpu;
@@ -1155,6 +1234,7 @@ struct test {
             "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
             "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
             "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
+            "defrag_thold",
             "use_mmap",     "embeddings",   "no_op_offload",   "n_prompt",       "n_gen",      "n_depth",      "test_time",
             "avg_ns",       "stddev_ns",    "avg_ts",         "stddev_ts",
         };
@@ -1174,7 +1254,7 @@ struct test {
             field == "use_mmap" || field == "embeddings") {
             return BOOL;
         }
-        if (field == "avg_ts" || field == "stddev_ts") {
+        if (field == "avg_ts" || field == "stddev_ts" || field == "defrag_thold") {
             return FLOAT;
         }
         return STRING;
@@ -1241,6 +1321,7 @@ struct test {
                                             std::to_string(flash_attn),
                                             tensor_split_str,
                                             tensor_buft_overrides_str,
+                                            std::to_string(defrag_thold),
                                             std::to_string(use_mmap),
                                             std::to_string(embeddings),
                                             std::to_string(no_op_offload),
@@ -1507,6 +1588,9 @@ struct markdown_printer : public printer {
         if (params.type_v.size() > 1 || params.type_v != cmd_params_defaults.type_v) {
             fields.emplace_back("type_v");
         }
+        if (params.defrag_thold.size() > 1 || params.defrag_thold != cmd_params_defaults.defrag_thold) {
+            fields.emplace_back("defrag_thold");
+        }
         if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
             fields.emplace_back("main_gpu");
         }

tools/mtmd/clip.cpp

@@ -1909,16 +1909,20 @@ struct clip_model_loader {
                     } break;
                 case PROJECTOR_TYPE_QWEN2VL:
                     {
-                        // max image size = sqrt(max_pixels)
-                        // https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/blob/main/preprocessor_config.json
-                        hparams.image_size = 3584;
+                        // max image size = sqrt(max_pixels) = 3584
+                        // ref: https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/blob/main/preprocessor_config.json
+                        // however, the model use unreasonable memory past 1024 size, we force it to 1024 otherwise it's unusable
+                        // ref: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/discussions/10
+                        hparams.image_size = 1024;
                         hparams.warmup_image_size = hparams.patch_size * 8;
                     } break;
                 case PROJECTOR_TYPE_QWEN25VL:
                     {
                         // max image size = sqrt(max_pixels)
                         // https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/blob/main/preprocessor_config.json
-                        hparams.image_size = 3584;
+                        // however, the model use unreasonable memory past 1024 size, we force it to 1024 otherwise it's unusable
+                        // ref: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/discussions/10
+                        hparams.image_size = 1024;
                         hparams.warmup_image_size = hparams.patch_size * 8;
                         get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern);
                     } break;

tools/server/utils.hpp

@@ -644,7 +644,7 @@ static json oaicompat_completion_params_parse(
     }
     for (auto & msg : messages) {
         json & content = msg.at("content");
-        if (content.is_string()) {
+        if (content.is_string() || content.is_null()) {
             continue;
         }