llama : Add Gemma 3 support (+ experimental vision capability) (#12343)

* llama : Add Gemma 3 text-only support

* fix python coding style

* fix compile on ubuntu

* python: fix style

* fix ubuntu compile

* fix build on ubuntu (again)

* fix ubuntu build, finally

* clip : Experimental support for Gemma 3 vision (#12344)

* clip : Experimental support for Gemma 3 vision

* fix build

* PRId64

Makefile

@@ -836,7 +836,7 @@ ifdef GGML_MUSA
 	else
 		MUSA_PATH ?= /opt/musa
 	endif
-	MUSA_ARCHITECTURES ?= 21;22
+	MUSA_ARCHITECTURES ?= 21;22;31
 
 	MK_CPPFLAGS += -DGGML_USE_MUSA -DGGML_USE_CUDA
 	MK_LDFLAGS += -L$(MUSA_PATH)/lib -Wl,-rpath=$(MUSA_PATH)/lib

convert_hf_to_gguf.py

@@ -861,6 +861,9 @@ class Model:
                 for token_id, token_data in added_tokens_decoder.items():
                     token_id = int(token_id)
                     token: str = token_data["content"]
+                    if token_id >= vocab_size:
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
                     if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
                         if tokens[token_id] != token.encode("utf-8"):
                             logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token!r}')
@@ -3322,6 +3325,83 @@ class Gemma2Model(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration")
+class Gemma3Model(Model):
+    model_arch = gguf.MODEL_ARCH.GEMMA3
+    has_vision: bool = False
+
+    # we need to merge the text_config into the root level of hparams
+    def __init__(self, *args, **kwargs):
+        hparams = Model.load_hparams(kwargs["dir_model"])
+        if "text_config" in hparams:
+            hparams = {**hparams, **hparams["text_config"]}
+            kwargs["hparams"] = hparams
+        super().__init__(*args, **kwargs)
+        if "vision_config" in hparams:
+            logger.info("Has vision encoder, but it will be ignored")
+            self.has_vision = True
+
+    def write(self):
+        super().write()
+        if self.has_vision:
+            logger.info("NOTE: this script only convert the language model to GGUF")
+            logger.info("      for the vision model, please use gemma3_convert_encoder_to_gguf.py")
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+        self.gguf_writer.add_add_space_prefix(False)
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+
+        # some default values are not specified in the hparams
+        self.gguf_writer.add_context_length(hparams.get("max_position_embeddings", 131072))
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 8))
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-6))
+        self.gguf_writer.add_key_length(hparams.get("head_dim", 256))
+        self.gguf_writer.add_value_length(hparams.get("head_dim", 256))
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1_000_000.0)) # for global layers
+        # both attn_logit_softcapping and final_logit_softcapping are removed in Gemma3
+        assert hparams.get("attn_logit_softcapping") is None
+        assert hparams.get("final_logit_softcapping") is None
+        self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+        self.gguf_writer.add_head_count_kv(hparams.get("num_key_value_heads", 4))
+        if hparams.get("rope_scaling") is not None:
+            assert hparams["rope_scaling"]["rope_type"] == "linear"
+            # important: this rope_scaling is only applied for global layers, and not used by 1B model
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        if name.startswith("language_model."):
+            name = name.replace("language_model.", "")
+        elif name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \
+                or name.startswith("multimodal_projector.") or name.startswith("vision_model."): # this is for old HF model, should be removed later
+            # ignore vision tensors
+            return []
+
+        # remove OOV (out-of-vocabulary) rows in token_embd
+        if "embed_tokens.weight" in name:
+            vocab = self._create_vocab_sentencepiece()
+            tokens = vocab[0]
+            data_torch = data_torch[:len(tokens)]
+
+        # ref code in Gemma3RMSNorm
+        # output = output * (1.0 + self.weight.float())
+        if name.endswith("norm.weight"):
+            data_torch = data_torch + 1
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @Model.register("Starcoder2ForCausalLM")
 class StarCoder2Model(Model):
     model_arch = gguf.MODEL_ARCH.STARCODER2

docs/build.md

@@ -197,29 +197,53 @@ The following compilation options are also available to tweak performance:
 
 ## MUSA
 
-This provides GPU acceleration using the MUSA cores of your Moore Threads MTT GPU. Make sure to have the MUSA SDK installed. You can download it from here: [MUSA SDK](https://developer.mthreads.com/sdk/download/musa).
+This provides GPU acceleration using a Moore Threads GPU. Make sure to have the [MUSA SDK](https://developer.mthreads.com/musa/musa-sdk) installed.
 
-- Using `CMake`:
+#### Download directly from Moore Threads
 
-  ```bash
-  cmake -B build -DGGML_MUSA=ON
-  cmake --build build --config Release
+You may find the official downloads here: [Moore Threads developer site](https://developer.mthreads.com/sdk/download/musa).
+
+### Compilation
+
+```bash
+cmake -B build -DGGML_MUSA=ON
+cmake --build build --config Release
+```
+
+#### Override Compute Capability Specifications
+
+By default, all supported compute capabilities are enabled. To customize this behavior, you can specify the `MUSA_ARCHITECTURES` option in the CMake command:
+
+```bash
+cmake -B build -DGGML_MUSA=ON -DMUSA_ARCHITECTURES="21"
+```
+
+This configuration enables only compute capability `2.1` (MTT S80) during compilation, which can help reduce compilation time.
+
+#### Compilation options
+
+Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet.
+
+- For static builds, add `-DBUILD_SHARED_LIBS=OFF` and `-DCMAKE_POSITION_INDEPENDENT_CODE=ON`:
   ```
-
-  For static build:
-
-  ```bash
   cmake -B build -DGGML_MUSA=ON \
     -DBUILD_SHARED_LIBS=OFF -DCMAKE_POSITION_INDEPENDENT_CODE=ON
   cmake --build build --config Release
   ```
 
-The environment variable [`MUSA_VISIBLE_DEVICES`](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) can be used to specify which GPU(s) will be used.
+### Runtime MUSA environmental variables
+
+You may set the [musa environmental variables](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) at runtime.
+
+```bash
+# Use `MUSA_VISIBLE_DEVICES` to hide the first compute device.
+MUSA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.gguf
+```
+
+### Unified Memory
 
 The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted.
 
-Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet.
-
 ## HIP
 
 This provides GPU acceleration on HIP-supported AMD GPUs.

examples/llava/CMakeLists.txt

@@ -51,6 +51,13 @@ install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
+set(TARGET llama-gemma3-cli)
+add_executable(${TARGET} gemma3-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-gemma3-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
+
 set(TARGET llama-llava-clip-quantize-cli)
 add_executable(${TARGET} clip-quantize-cli.cpp)
 set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-clip-quantize-cli)

examples/llava/README-gemma3.md

@@ -0,0 +1,30 @@
+# Gemma 3 vision
+
+> [!IMPORTANT]
+>
+> This is very experimental, only used for demo purpose.
+
+## How to get mmproj.gguf?
+
+```bash
+cd gemma-3-4b-it
+python ../llama.cpp/examples/llava/gemma3_convert_encoder_to_gguf.py .
+
+# output file is mmproj.gguf
+```
+
+## How to run it?
+
+What you need:
+- The text model GGUF, can be converted using `convert_hf_to_gguf.py`
+- The mmproj file from step above
+- An image file
+
+```bash
+# build
+cmake -B build
+cmake --build build --target llama-gemma3-cli
+
+# run it
+./build/bin/llama-gemma3-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
+```

examples/llava/clip.cpp

@@ -4,31 +4,12 @@
 // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
 #include "clip.h"
 #include "ggml.h"
+#include "ggml-cpp.h"
 #include "ggml-cpu.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 #include "gguf.h"
 
-//#ifdef GGML_USE_CUDA
-//#include "ggml-cuda.h"
-//#endif
-//
-//#ifdef GGML_USE_SYCL
-//#include "ggml-sycl.h"
-//#endif
-//
-//#ifdef GGML_USE_METAL
-//#include "ggml-metal.h"
-//#endif
-//
-//#ifdef GGML_USE_CANN
-//#include "ggml-cann.h"
-//#endif
-//
-//#ifdef GGML_USE_VULKAN
-//#include "ggml-vulkan.h"
-//#endif
-
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
@@ -155,6 +136,8 @@ static std::string format(const char * fmt, ...) {
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
+#define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
+#define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
 
 #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
 #define TN_MINICPMV_QUERY "resampler.query"
@@ -181,6 +164,7 @@ enum projector_type {
     PROJECTOR_TYPE_RESAMPLER,
     PROJECTOR_TYPE_GLM_EDGE,
     PROJECTOR_TYPE_MERGER,
+    PROJECTOR_TYPE_GEMMA3,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -191,6 +175,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_RESAMPLER, "resampler"},
     { PROJECTOR_TYPE_GLM_EDGE, "adapter"},
     { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"},
+    { PROJECTOR_TYPE_GEMMA3, "gemma3"},
 };
 
 
@@ -317,7 +302,7 @@ static projector_type clip_projector_type_from_string(const std::string & name)
             return kv.first;
         }
     }
-    return PROJECTOR_TYPE_UNKNOWN;
+    throw std::runtime_error(format("Unknown projector type: %s", name.c_str()));
 }
 
 #ifdef CLIP_DEBUG_FUNCTIONS
@@ -574,6 +559,10 @@ struct clip_vision_model {
     struct ggml_tensor * mm_model_ln_kv_b;
     struct ggml_tensor * mm_model_ln_post_w;
     struct ggml_tensor * mm_model_ln_post_b;
+
+    // gemma3
+    struct ggml_tensor * mm_input_proj_w;
+    struct ggml_tensor * mm_soft_emb_norm_w;
 };
 
 struct clip_ctx {
@@ -588,7 +577,7 @@ struct clip_ctx {
     struct clip_vision_model vision_model;
     projector_type proj_type = PROJECTOR_TYPE_MLP;
 
-    int32_t max_feature_layer;
+    int32_t max_feature_layer; // unused in newer models like gemma3
     float image_mean[3];
     float image_std[3];
     bool use_gelu = false;
@@ -600,21 +589,209 @@ struct clip_ctx {
     bool has_post_norm = false;
     bool has_patch_bias = false;
 
-    struct gguf_context * ctx_gguf;
-    struct ggml_context * ctx_data;
+    struct gguf_context * ctx_gguf = nullptr;
+    struct ggml_context * ctx_data = nullptr;
 
     std::vector<uint8_t> buf_compute_meta;
 
-    // memory buffers to evaluate the model
-    ggml_backend_buffer_t params_buffer  = NULL;
+    std::vector<ggml_backend_t> backend_ptrs;
+    std::vector<ggml_backend_buffer_type_t> backend_buft;
 
-    ggml_backend_t backend       = NULL;
-    ggml_gallocr_t compute_alloc = NULL;
+    ggml_backend_t backend     = nullptr;
+    ggml_backend_t backend_cpu = nullptr;
+    ggml_backend_buffer_t buf  = nullptr;
 
-    struct clip_image_size * load_image_size;
+    ggml_backend_sched_ptr sched;
+
+    struct clip_image_size * load_image_size = nullptr;
+
+    clip_ctx(clip_context_params & ctx_params) {
+        backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
+        backend     = ctx_params.use_gpu
+                        ? ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr)
+                        : nullptr;
+
+        if (backend) {
+            LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend));
+            backend_ptrs.push_back(backend);
+            backend_buft.push_back(ggml_backend_get_default_buffer_type(backend));
+        } else {
+            backend = backend_cpu;
+            LOG_INF("%s: CLIP using CPU backend\n", __func__);
+        }
+
+        backend_ptrs.push_back(backend_cpu);
+        backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu));
+
+        sched.reset(
+            ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false)
+        );
+    }
+
+    ~clip_ctx() {
+        ggml_free(ctx_data);
+        gguf_free(ctx_gguf);
+        ggml_backend_buffer_free(buf);
+        ggml_backend_free(backend);
+        if (backend_cpu != backend) {
+            ggml_backend_free(backend_cpu);
+        }
+    }
 };
 
-static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
+static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
+    const auto & model = ctx->vision_model;
+    const auto & hparams = model.hparams;
+
+    const int image_size = hparams.image_size;
+    int image_size_width  = image_size;
+    int image_size_height = image_size;
+
+    const int patch_size           = hparams.patch_size;
+    const int num_patches          = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int hidden_size          = hparams.hidden_size;
+    const int n_head               = hparams.n_head;
+    const int d_head               = hidden_size / n_head;
+    const int n_layer              = hparams.n_layer;
+    const float eps                = hparams.eps;
+
+    GGML_ASSERT(imgs->size == 1); // batch_size == 1
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ ctx->buf_compute_meta.size(),
+        /*.mem_buffer =*/ ctx->buf_compute_meta.data(),
+        /*.no_alloc   =*/ true,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    // input raw
+    struct ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3);
+    ggml_set_name(inp_raw, "inp_raw");
+    ggml_set_input(inp_raw);
+
+    struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+    inp = ggml_reshape_2d(ctx0, inp, num_patches, hidden_size);
+    inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
+    inp = ggml_add(ctx0, inp, model.patch_bias);
+
+    // position embeddings
+    struct ggml_tensor * embeddings = ggml_add(ctx0, inp, model.position_embeddings);
+
+    // loop over layers
+    for (int il = 0; il < n_layer; il++) {
+        struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
+
+        // layernorm1
+        {
+            cur = ggml_norm(ctx0, cur, eps);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_1_w), model.layers[il].ln_1_b);
+        }
+
+        // self-attention
+        {
+
+            struct ggml_tensor * Q =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b);
+
+            Q = ggml_reshape_3d(ctx0, Q, d_head, n_head, num_patches);
+            Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+
+            struct ggml_tensor * K =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b);
+
+            K = ggml_reshape_3d(ctx0, K, d_head, n_head, num_patches);
+            K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+
+            struct ggml_tensor * V =
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].v_w, cur), model.layers[il].v_b);
+
+            V = ggml_reshape_3d(ctx0, V, d_head, n_head, num_patches);
+            V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            KQ = ggml_scale_inplace(ctx0, KQ, 1.0f / sqrtf((float)d_head));
+            KQ = ggml_soft_max_inplace(ctx0, KQ);
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
+            KQV = ggml_reshape_3d(ctx0, KQV, d_head, num_patches, n_head);
+            KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            cur = ggml_cont_2d(ctx0, KQV, hidden_size, num_patches);
+        }
+
+        // attention output
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].o_w, cur), model.layers[il].o_b);
+
+        // re-add the layer input, e.g., residual
+        cur = ggml_add(ctx0, cur, embeddings);
+
+        embeddings = cur; // embeddings = residual, cur = hidden_states
+
+        // layernorm2
+        {
+            cur = ggml_norm(ctx0, cur, eps);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_2_w), model.layers[il].ln_2_b);
+        }
+
+        cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_i_b);
+
+        // siglip uses gelu
+        cur = ggml_gelu(ctx0, cur);
+
+        cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_o_b);
+
+        // residual 2
+        cur = ggml_add(ctx0, embeddings, cur);
+
+        embeddings = cur;
+    }
+
+    // post-layernorm
+    if (ctx->has_post_norm) {
+        embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "post_ln");
+
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b);
+    }
+
+    if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
+        const int batch_size = 1;
+        const int mm_tokens_per_image = 256; // default value for gemma3
+        const int tokens_per_side = sqrt(mm_tokens_per_image);
+        const int patches_per_image = sqrt(num_patches);
+        const int kernel_size = patches_per_image / tokens_per_side;
+
+        embeddings = ggml_cont(ctx0, ggml_transpose(ctx0, embeddings));
+        embeddings = ggml_reshape_4d(ctx0, embeddings, patches_per_image, patches_per_image, hidden_size, batch_size);
+
+        // doing a pool2d to reduce the number of output tokens to 256
+        embeddings = ggml_pool_2d(ctx0, embeddings, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0);
+        embeddings = ggml_reshape_3d(ctx0, embeddings, embeddings->ne[0] * embeddings->ne[0], hidden_size, batch_size);
+        embeddings = ggml_cont(ctx0, ggml_transpose(ctx0, embeddings));
+
+        // apply norm before projection
+        embeddings = ggml_rms_norm(ctx0, embeddings, eps);
+        embeddings = ggml_mul(ctx0, embeddings, model.mm_soft_emb_norm_w);
+
+        // apply projection
+        embeddings = ggml_mul_mat(ctx0,
+            ggml_cont(ctx0, ggml_transpose(ctx0, model.mm_input_proj_w)),
+            embeddings);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
     if (!ctx->has_vision_encoder) {
         LOG_ERR("This gguf file seems to have no vision encoder\n");
         return nullptr;
@@ -1160,7 +1337,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         } else {
             GGML_ABORT("fatel error");
         }
-    } else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
+    }
+    else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
         embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size);
 
         embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
@@ -1182,8 +1360,25 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     return gf;
 }
 
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
+    if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
+        return clip_image_build_graph_siglip(ctx, imgs);
+    } else {
+        // TODO: we should have one build_* function per model
+        return clip_image_build_graph_legacy(ctx, imgs, load_image_size, is_inf);
+    }
+}
+
 // read and create ggml_context containing the tensors and their data
 struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
+    return clip_init(fname, clip_context_params{
+        /* use_gpu */   true,
+        /* verbosity */ verbosity,
+    });
+}
+
+struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params) {
+    int verbosity = ctx_params.verbosity;
     struct ggml_context * meta = NULL;
 
     struct gguf_init_params params = {
@@ -1277,7 +1472,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-    clip_ctx * new_clip = new clip_ctx{};
+    clip_ctx * new_clip = new clip_ctx(ctx_params);
 
     // update projector type
     {
@@ -1296,36 +1491,6 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-//#ifdef GGML_USE_CUDA
-//    new_clip->backend = ggml_backend_cuda_init(0);
-//    LOG_INF("%s: CLIP using CUDA backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_METAL
-//    new_clip->backend = ggml_backend_metal_init();
-//    LOG_INF("%s: CLIP using Metal backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_CANN
-//    new_clip->backend = ggml_backend_cann_init(0);
-//    LOG_INF("%s: CLIP using CANN backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_VULKAN
-//    new_clip->backend = ggml_backend_vk_init(0);
-//    LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
-//#endif
-//
-//#ifdef GGML_USE_SYCL
-//    new_clip->backend = ggml_backend_sycl_init(0);
-//    LOG_INF("%s: CLIP using SYCL backend\n", __func__);
-//#endif
-
-    if (!new_clip->backend) {
-        new_clip->backend = ggml_backend_cpu_init();
-        LOG_INF("%s: CLIP using CPU backend\n", __func__);
-    }
-
     // model size and capabilities
     {
         int idx = get_key_idx(ctx, KEY_HAS_TEXT_ENC);
@@ -1363,8 +1528,12 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         GGML_ASSERT(new_clip->has_vision_encoder);
         GGML_ASSERT(!new_clip->has_text_encoder);
 
-        idx = get_key_idx(ctx, KEY_USE_GELU);
-        new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
+        try {
+            idx = get_key_idx(ctx, KEY_USE_GELU);
+            new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
+        } catch (std::runtime_error & /*e*/) {
+            new_clip->use_gelu = false;
+        }
 
         try {
             idx = get_key_idx(ctx, KEY_USE_SILU);
@@ -1421,7 +1590,9 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         // alloc memory and offload data
-        new_clip->params_buffer = ggml_backend_alloc_ctx_tensors(new_clip->ctx_data, new_clip->backend);
+        ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(new_clip->backend);
+        new_clip->buf = ggml_backend_alloc_ctx_tensors_from_buft(new_clip->ctx_data, buft);
+        ggml_backend_buffer_set_usage(new_clip->buf, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
         for (int i = 0; i < n_tensors; ++i) {
             const char * name = gguf_get_tensor_name(ctx, i);
             struct ggml_tensor * cur = ggml_get_tensor(new_clip->ctx_data, name);
@@ -1434,7 +1605,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
                 return nullptr;
             }
             int num_bytes = ggml_nbytes(cur);
-            if (ggml_backend_buffer_is_host(new_clip->params_buffer)) {
+            if (ggml_backend_buft_is_host(buft)) {
                 // for the CPU and Metal backend, we can read directly into the tensor
                 fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
             } else {
@@ -1570,11 +1741,17 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         try {
-            vision_model.patch_embeddings_0    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+            vision_model.patch_embeddings_0 = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+        } catch(const std::exception& /*e*/) {
+            vision_model.patch_embeddings_0 = nullptr;
+        }
+
+        try {
             vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
         } catch(const std::exception& /*e*/) {
-            LOG_ERR("%s: failed to load vision model tensors\n", __func__);
+            vision_model.position_embeddings = nullptr;
         }
+
         try {
             vision_model.patch_embeddings_1    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD_1);
         } catch(const std::exception& /*e*/) {
@@ -1685,6 +1862,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
             vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
         }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_GEMMA3) {
+            vision_model.mm_input_proj_w    = get_tensor(new_clip->ctx_data, TN_MM_INP_PROJ);
+            vision_model.mm_soft_emb_norm_w = get_tensor(new_clip->ctx_data, TN_MM_SOFT_EMB_N);
+        }
         else {
             std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
             throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
@@ -1720,14 +1901,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     // measure mem requirement and allocate
     {
         new_clip->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
-        new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
         clip_image_f32_batch batch;
         batch.size = 1;
         batch.data = nullptr;
         ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
-        ggml_gallocr_reserve(new_clip->compute_alloc, gf);
-        size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
-        LOG_INF("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
+        ggml_backend_sched_reserve(new_clip->sched.get(), gf);
+        for (size_t i = 0; i < new_clip->backend_ptrs.size(); ++i) {
+            ggml_backend_t backend = new_clip->backend_ptrs[i];
+            ggml_backend_buffer_type_t buft = new_clip->backend_buft[i];
+            size_t size = ggml_backend_sched_get_buffer_size(new_clip->sched.get(), backend);
+            if (size > 1) {
+                LOG_INF("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
+                        ggml_backend_buft_name(buft),
+                        size / 1024.0 / 1024.0);
+            }
+        }
     }
 
     return new_clip;
@@ -2219,7 +2407,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
         return true;
     }
 
-    if (ctx->has_glm_projector) {
+    if (ctx->has_glm_projector || ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
         res_imgs->size = 1;
         res_imgs->data = new clip_image_f32[res_imgs->size];
         clip_image_u8 resized_image;
@@ -2408,12 +2596,6 @@ ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
 }
 
 void clip_free(clip_ctx * ctx) {
-    ggml_free(ctx->ctx_data);
-    gguf_free(ctx->ctx_gguf);
-
-    ggml_backend_buffer_free(ctx->params_buffer);
-    ggml_backend_free(ctx->backend);
-    ggml_gallocr_free(ctx->compute_alloc);
     delete ctx;
 }
 
@@ -2609,8 +2791,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     }
 
     // build the inference graph
+    ggml_backend_sched_reset(ctx->sched.get());
     ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true);
-    ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
+    ggml_backend_sched_alloc_graph(ctx->sched.get(), gf);
 
     // set inputs
     const auto & model = ctx->vision_model;
@@ -2749,6 +2932,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
             free(positions_data);
         }
+        else if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
+            // do nothing
+        }
         else {
             struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
 
@@ -2775,11 +2961,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         }
     }
 
-    if (ggml_backend_is_cpu(ctx->backend)) {
-        ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
-    }
+    ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads);
 
-    ggml_backend_graph_compute(ctx->backend, gf);
+    auto status = ggml_backend_sched_graph_compute(ctx->sched.get(), gf);
+    if (status != GGML_STATUS_SUCCESS) {
+        LOG_ERR("%s: ggml_backend_sched_graph_compute failed with error %d\n", __func__, status);
+        return false;
+    }
 
     // the last node is the embedding tensor
     struct ggml_tensor * embeddings = ggml_graph_node(gf, -1);
@@ -2959,6 +3147,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
         return ctx->vision_model.mm_1_b->ne[0];
     }
+    if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) {
+        return ctx->vision_model.mm_input_proj_w->ne[0];
+    }
 
     std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
     throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));

examples/llava/clip.h

@@ -39,8 +39,15 @@ struct clip_image_f32_batch {
     size_t size;
 };
 
-CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
-CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
+struct clip_context_params {
+    bool use_gpu;
+    int verbosity;
+};
+
+// deprecated, use clip_init
+CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
+
+CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params);
 
 CLIP_API void clip_free(struct clip_ctx * ctx);
 

examples/llava/gemma3-cli.cpp

@@ -0,0 +1,341 @@
+#include "arg.h"
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+#include "clip.h"
+#include "stb_image.h"
+#include "llama.h"
+#include "ggml.h"
+#include "console.h"
+
+#include <vector>
+#include <limits.h>
+#include <inttypes.h>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+static bool g_is_generating = false;
+
+/**
+ * Please note that this is NOT a production-ready stuff.
+ * It is a playground for trying Gemma 3 vision capabilities.
+ * For contributors: please keep this code simple and easy to understand.
+ */
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    LOG(
+        "Experimental CLI for using Gemma 3 vision model\n\n"
+        "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n"
+        "  -m and --mmproj are required\n"
+        "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n",
+        argv[0]
+    );
+}
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void sigint_handler(int signo) {
+    if (signo == SIGINT) {
+        if (g_is_generating) {
+            g_is_generating = false;
+        } else {
+            console::cleanup();
+            LOG("\nInterrupted by user\n");
+            _exit(130);
+        }
+    }
+}
+#endif
+
+struct gemma3_context {
+    struct clip_ctx    * ctx_clip = NULL;
+    common_init_result   llama_init;
+
+    llama_model       * model;
+    llama_context     * lctx;
+    const llama_vocab * vocab;
+    llama_batch         batch;
+
+    int n_threads    = 1;
+    llama_pos n_past = 0;
+
+    gemma3_context(common_params & params) : llama_init(common_init_from_params(params)) {
+        model = llama_init.model.get();
+        lctx = llama_init.context.get();
+        vocab = llama_model_get_vocab(model);
+        n_threads = params.cpuparams.n_threads;
+        batch = llama_batch_init(params.n_batch, 0, 1);
+        init_clip_model(params);
+    }
+
+    void init_clip_model(common_params & params) {
+        const char * clip_path = params.mmproj.c_str();
+        ctx_clip = clip_model_load(clip_path, params.verbosity > 1);
+    }
+
+    ~gemma3_context() {
+        clip_free(ctx_clip);
+    }
+};
+
+struct decode_embd_batch {
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id>   seq_id_0;
+    std::vector<llama_seq_id *> seq_ids;
+    std::vector<int8_t>         logits;
+    llama_batch batch;
+    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+        pos     .resize(n_tokens);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        seq_id_0[0] = seq_id;
+        seq_ids [n_tokens] = nullptr;
+        batch = {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ embd,
+            /*pos            =*/ pos.data(),
+            /*n_seq_id       =*/ n_seq_id.data(),
+            /*seq_id         =*/ seq_ids.data(),
+            /*logits         =*/ logits.data(),
+        };
+        for (int i = 0; i < n_tokens; i++) {
+            batch.pos     [i] = pos_0 + i;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+};
+
+static int eval_text(gemma3_context & ctx, std::string input, bool logits_last = false) {
+    llama_tokens tokens = common_tokenize(ctx.lctx, input, false, true);
+    common_batch_clear(ctx.batch);
+    for (llama_token & t : tokens) {
+        common_batch_add(ctx.batch, t, ctx.n_past++, {0}, false);
+    }
+    if (logits_last) {
+        ctx.batch.logits[ctx.batch.n_tokens - 1] = true;
+    }
+    // LOG("eval_text (n_tokens = %d): %s\n", (int)tokens.size(), input.c_str());
+    if (llama_decode(ctx.lctx, ctx.batch)) {
+        LOG_ERR("Failed to decode text\n");
+        return 1;
+    }
+    return 0;
+}
+
+static int eval_image(gemma3_context & ctx, std::string & fname) {
+    std::vector<float> image_embd_v;
+    int n_embd = llama_model_n_embd(ctx.model);
+    int n_tokens = 256;
+    image_embd_v.resize(n_tokens * n_embd);
+
+    bool ok;
+    struct clip_image_u8 * img_u8 = clip_image_u8_init();
+    ok = clip_image_load_from_file(fname.c_str(), img_u8);
+    if (!ok) {
+        LOG_ERR("Unable to load image %s\n", fname.c_str());
+        clip_image_u8_free(img_u8);
+        return 2; // non-fatal error
+    }
+
+    clip_image_f32_batch batch_f32;
+    ok = clip_image_preprocess(ctx.ctx_clip, img_u8, &batch_f32);
+    if (!ok) {
+        LOG_ERR("Unable to preprocess image\n");
+        clip_image_f32_batch_free(&batch_f32);
+        clip_image_u8_free(img_u8);
+        return 1;
+    }
+
+    int64_t t0 = ggml_time_ms();
+    LOG("Encoding image %s\n", fname.c_str());
+    ok = clip_image_batch_encode(ctx.ctx_clip, ctx.n_threads, &batch_f32, image_embd_v.data());
+    if (!ok) {
+        LOG_ERR("Unable to encode image\n");
+        clip_image_f32_batch_free(&batch_f32);
+        clip_image_u8_free(img_u8);
+        return 1;
+    }
+    LOG("Image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
+
+    clip_image_f32_batch_free(&batch_f32);
+    clip_image_u8_free(img_u8);
+
+    // decode image embeddings
+    int64_t t1 = ggml_time_ms();
+    eval_text(ctx, "<start_of_image>");
+    llama_set_causal_attn(ctx.lctx, false);
+    decode_embd_batch batch_img(image_embd_v.data(), n_tokens, ctx.n_past, 0);
+    if (llama_decode(ctx.lctx, batch_img.batch)) {
+        LOG_ERR("failed to decode image\n");
+        return 1;
+    }
+    ctx.n_past += n_tokens;
+    llama_set_causal_attn(ctx.lctx, true);
+    eval_text(ctx, "<end_of_image>");
+    LOG("Image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1);
+    return 0;
+}
+
+static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_predict) {
+    for (int i = 0; i < n_predict; i++) {
+        if (i > n_predict || !g_is_generating) {
+            printf("\n");
+            break;
+        }
+
+        llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1);
+        common_sampler_accept(smpl, token_id, true);
+
+        if (llama_vocab_is_eog(ctx.vocab, token_id)) {
+            printf("\n");
+            break; // end of generation
+        }
+
+        printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
+        fflush(stdout);
+
+        // eval the token
+        common_batch_clear(ctx.batch);
+        common_batch_add(ctx.batch, token_id, ctx.n_past++, {0}, true);
+        if (llama_decode(ctx.lctx, ctx.batch)) {
+            LOG_ERR("failed to decode token\n");
+            return 1;
+        }
+    }
+    return 0;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    common_params params;
+    params.sampling.temp = 0.2; // lower temp by default for better quality
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
+        return 1;
+    }
+
+    common_init();
+
+    if (params.mmproj.empty()) {
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+    gemma3_context ctx(params);
+    printf("%s: %s\n", __func__, params.model.c_str());
+
+    bool is_single_turn = !params.prompt.empty() && !params.image.empty();
+
+    struct common_sampler * smpl = common_sampler_init(ctx.model, params.sampling);
+    int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict;
+
+    // ctrl+C handling
+    {
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = sigint_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+    }
+
+    if (eval_text(ctx, "<bos>")) {
+        return 1;
+    }
+
+    if (is_single_turn) {
+        g_is_generating = true;
+        if (eval_text(ctx, "<start_of_turn>user\n")) {
+            return 1;
+        }
+        for (auto & fname : params.image) {
+            if (eval_image(ctx, fname)) {
+                return 1;
+            }
+        }
+        if (eval_text(ctx, params.prompt + "<end_of_turn><start_of_turn>model\n", true)) {
+            return 1;
+        }
+        if (generate_response(ctx, smpl, n_predict)) {
+            return 1;
+        }
+
+    } else {
+        LOG("\n Running in chat mode, available commands:");
+        LOG("\n   /image <path>    load an image");
+        LOG("\n   /clear           clear the chat history");
+        LOG("\n   /quit or /exit   exit the program");
+        LOG("\n");
+
+        if (eval_text(ctx, "<start_of_turn>user\n")) {
+            return 1;
+        }
+
+        while (true) {
+            g_is_generating = false;
+            LOG("\n> ");
+            console::set_display(console::user_input);
+            std::string line;
+            console::readline(line, false);
+            console::set_display(console::reset);
+            line = string_strip(line);
+            if (line.empty()) {
+                continue;
+            }
+            if (line == "/quit" || line == "/exit") {
+                break;
+            }
+            if (line == "/clear") {
+                ctx.n_past = 0;
+                llama_kv_cache_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS
+                LOG("Chat history cleared\n\n");
+                continue;
+            }
+            g_is_generating = true;
+            if (line.find("/image") == 0) {
+                std::string image = line.substr(7);
+                int res = eval_image(ctx, image);
+                if (res == 2) {
+                    continue; // image not found
+                }
+                if (res) {
+                    return 1;
+                }
+                continue;
+            }
+            if (eval_text(ctx, line + "<end_of_turn><start_of_turn>model\n", true)) {
+                return 1;
+            }
+            if (generate_response(ctx, smpl, n_predict)) {
+                return 1;
+            }
+            if (eval_text(ctx, "<end_of_turn><start_of_turn>user\n")) {
+                return 1;
+            }
+        }
+    }
+
+    return 0;
+}

examples/llava/gemma3_convert_encoder_to_gguf.py

@@ -0,0 +1,307 @@
+import gguf
+import argparse
+import logging
+import sys
+import torch
+import json
+import os
+import numpy as np
+from typing import cast, ContextManager, Any, Iterator
+from pathlib import Path
+from torch import Tensor
+
+logger = logging.getLogger("gemma3-mmproj")
+
+
+# (copied from convert_hf_to_gguf.py)
+# tree of lazy tensors
+class LazyTorchTensor(gguf.LazyBase):
+    _tensor_type = torch.Tensor
+    # to keep the type-checker happy
+    dtype: torch.dtype
+    shape: torch.Size
+
+    # only used when converting a torch.Tensor to a np.ndarray
+    _dtype_map: dict[torch.dtype, type] = {
+        torch.float16: np.float16,
+        torch.float32: np.float32,
+    }
+
+    # used for safetensors slices
+    # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046
+    # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734
+    _dtype_str_map: dict[str, torch.dtype] = {
+        "F64": torch.float64,
+        "F32": torch.float32,
+        "BF16": torch.bfloat16,
+        "F16": torch.float16,
+        # "U64": torch.uint64,
+        "I64": torch.int64,
+        # "U32": torch.uint32,
+        "I32": torch.int32,
+        # "U16": torch.uint16,
+        "I16": torch.int16,
+        "U8": torch.uint8,
+        "I8": torch.int8,
+        "BOOL": torch.bool,
+        "F8_E4M3": torch.float8_e4m3fn,
+        "F8_E5M2": torch.float8_e5m2,
+    }
+
+    def numpy(self) -> gguf.LazyNumpyTensor:
+        dtype = self._dtype_map[self.dtype]
+        return gguf.LazyNumpyTensor(
+            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape),
+            args=(self,),
+            func=(lambda s: s.numpy())
+        )
+
+    @classmethod
+    def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: tuple[int, ...]) -> Tensor:
+        return torch.empty(size=shape, dtype=dtype, device="meta")
+
+    @classmethod
+    def from_safetensors_slice(cls, st_slice: Any) -> Tensor:
+        dtype = cls._dtype_str_map[st_slice.get_dtype()]
+        shape: tuple[int, ...] = tuple(st_slice.get_shape())
+        lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:])
+        return cast(torch.Tensor, lazy)
+
+    @classmethod
+    def __torch_function__(cls, func, types, args=(), kwargs=None):
+        del types  # unused
+
+        if kwargs is None:
+            kwargs = {}
+
+        if func is torch.Tensor.numpy:
+            return args[0].numpy()
+
+        return cls._wrap_fn(func)(*args, **kwargs)
+
+
+class Gemma3VisionTower:
+    hparams: dict
+    gguf_writer: gguf.GGUFWriter
+    fname_out: Path
+    ftype: gguf.LlamaFileType
+
+    @staticmethod
+    def load_hparams(dir_model: Path):
+        with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+            return json.load(f)
+
+    @staticmethod
+    def get_model_part_names(dir_model: Path, prefix: str, suffix: str) -> list[str]:
+        part_names: list[str] = []
+        for filename in os.listdir(dir_model):
+            if filename.startswith(prefix) and filename.endswith(suffix):
+                part_names.append(filename)
+        part_names.sort()
+        return part_names
+
+    def __init__(self,
+                 dir_model: Path,
+                 fname_out: Path,
+                 ftype: gguf.LlamaFileType,
+                 is_big_endian: bool,):
+        hparams = Gemma3VisionTower.load_hparams(dir_model)
+        self.hparams = hparams
+        self.fname_out = fname_out
+        self.ftype = ftype
+        endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
+        self.gguf_writer = gguf.GGUFWriter(path=None, arch="clip", endianess=endianess)
+
+        text_config = hparams["text_config"]
+        vision_config = hparams["vision_config"]
+
+        assert hparams["architectures"][0] == "Gemma3ForConditionalGeneration"
+        assert text_config is not None
+        assert vision_config is not None
+
+        self.gguf_writer.add_string ("clip.projector_type",              "gemma3")
+        self.gguf_writer.add_bool   ("clip.has_text_encoder",            False)
+        self.gguf_writer.add_bool   ("clip.has_vision_encoder",          True)
+        self.gguf_writer.add_bool   ("clip.has_llava_projector",         False) # legacy
+        self.gguf_writer.add_uint32 ("clip.vision.image_size",           vision_config["image_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.patch_size",           vision_config["patch_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.embedding_length",     vision_config["hidden_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.feed_forward_length",  vision_config["intermediate_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.projection_dim",       text_config["hidden_size"])
+        self.gguf_writer.add_uint32 ("clip.vision.block_count",          vision_config["num_hidden_layers"])
+        self.gguf_writer.add_uint32 ("clip.vision.attention.head_count", vision_config["num_attention_heads"])
+        self.gguf_writer.add_float32("clip.vision.attention.layer_norm_epsilon", vision_config.get("layer_norm_eps", 1e-6))
+        # default values taken from HF tranformers code
+        self.gguf_writer.add_array  ("clip.vision.image_mean", [0.5, 0.5, 0.5])
+        self.gguf_writer.add_array  ("clip.vision.image_std",  [0.5, 0.5, 0.5])
+        self.gguf_writer.add_bool   ("clip.use_gelu", True)
+
+        # load tensors
+        for name, data_torch in self.get_tensors(dir_model):
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+            self.add_tensor(name, data_torch)
+
+    def get_tensors(self, dir_model: Path) -> Iterator[tuple[str, Tensor]]:
+        part_names = Gemma3VisionTower.get_model_part_names(dir_model, "model", ".safetensors")
+        tensor_names_from_parts: set[str] = set()
+        for part_name in part_names:
+            logger.info(f"gguf: loading model part '{part_name}'")
+            from safetensors import safe_open
+            ctx = cast(ContextManager[Any], safe_open(dir_model / part_name, framework="pt", device="cpu"))
+            with ctx as model_part:
+                tensor_names_from_parts.update(model_part.keys())
+
+                for name in model_part.keys():
+                    data = model_part.get_slice(name)
+                    data = LazyTorchTensor.from_safetensors_slice(data)
+                    yield name, data
+
+    def add_tensor(self, name: str, data_torch: Tensor):
+        is_1d = len(data_torch.shape) == 1
+        is_embd = ".embeddings." in name
+        old_dtype = data_torch.dtype
+        can_quantize = not is_1d and not is_embd
+        data_qtype = gguf.GGMLQuantizationType.F32
+
+        # this is to support old checkpoint
+        # TODO: remove this when we have the final model
+        name = name.replace("vision_model.vision_model.", "vision_tower.vision_model.")
+        name = name.replace("multimodal_projector.", "multi_modal_projector.")
+
+        # filter only vision tensors
+        if not name.startswith("vision_tower.vision_model.") and not name.startswith("multi_modal_projector."):
+            return
+        # prefix
+        name = name.replace("vision_tower.vision_model.encoder.layers.", "v.blk.")
+        name = name.replace("vision_tower.vision_model.", "v.")
+        # projector and input embd
+        name = name.replace(".embeddings.patch_embedding.", ".patch_embd.")
+        name = name.replace(".embeddings.position_embedding.", ".position_embd.")
+        name = name.replace(
+            "multi_modal_projector.mm_input_projection_weight",
+            "mm.input_projection.weight"
+        )
+        name = name.replace(
+            "multi_modal_projector.mm_soft_emb_norm.weight",
+            "mm.soft_emb_norm.weight"
+        )
+        name = name.replace("post_layernorm.", "post_ln.")
+        # each block
+        name = name.replace(".self_attn.k_proj.", ".attn_k.")
+        name = name.replace(".self_attn.v_proj.", ".attn_v.")
+        name = name.replace(".self_attn.q_proj.", ".attn_q.")
+        name = name.replace(".self_attn.out_proj.", ".attn_out.")
+        name = name.replace(".layer_norm1.", ".ln1.")
+        name = name.replace(".layer_norm2.", ".ln2.")
+        name = name.replace(".mlp.fc1.", ".ffn_down.")
+        name = name.replace(".mlp.fc2.", ".ffn_up.")
+
+        if can_quantize:
+            if self.ftype == gguf.LlamaFileType.ALL_F32:
+                data_qtype = gguf.GGMLQuantizationType.F32
+            elif self.ftype == gguf.LlamaFileType.MOSTLY_F16:
+                data_qtype = gguf.GGMLQuantizationType.F16
+            elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                data_qtype = gguf.GGMLQuantizationType.BF16
+            elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
+                data_qtype = gguf.GGMLQuantizationType.Q8_0
+            else:
+                raise ValueError(f"Unsupported file type: {self.ftype}")
+
+        # corrent norm value ; only this "soft_emb_norm" need to be corrected as it's part of Gemma projector
+        # the other norm values are part of SigLIP model, and they are already correct
+        # ref code: Gemma3RMSNorm
+        if "soft_emb_norm.weight" in name:
+            logger.info(f"Correcting norm value for '{name}'")
+            data_torch = data_torch + 1
+
+        data = data_torch.numpy()
+
+        try:
+            data = gguf.quants.quantize(data, data_qtype)
+        except Exception as e:
+            logger.error(f"Error quantizing tensor '{name}': {e}, fallback to F16")
+            data_qtype = gguf.GGMLQuantizationType.F16
+            data = gguf.quants.quantize(data, data_qtype)
+
+        # reverse shape to make it similar to the internal ggml dimension order
+        shape_str = f"{{{', '.join(str(n) for n in reversed(data_torch.shape))}}}"
+        logger.info(f"{f'%-32s' % f'{name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
+
+        self.gguf_writer.add_tensor(name, data, raw_dtype=data_qtype)
+
+    def write(self):
+        self.gguf_writer.write_header_to_file(path=self.fname_out)
+        self.gguf_writer.write_kv_data_to_file()
+        self.gguf_writer.write_tensors_to_file(progress=True)
+        self.gguf_writer.close()
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Convert Gemma 3 vision tower safetensors to GGUF format",)
+    parser.add_argument(
+        "--outfile", type=Path, default="mmproj.gguf",
+        help="path to write to",
+    )
+    parser.add_argument(
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0"], default="f16",
+        help="output format",
+    )
+    parser.add_argument(
+        "--bigendian", action="store_true",
+        help="model is executed on big endian machine",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file",
+        nargs="?",
+    )
+    parser.add_argument(
+        "--verbose", action="store_true",
+        help="increase output verbosity",
+    )
+
+    args = parser.parse_args()
+    if args.model is None:
+        parser.error("the following arguments are required: model")
+    return args
+
+
+def main() -> None:
+    args = parse_args()
+
+    if args.verbose:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    dir_model = args.model
+
+    if not dir_model.is_dir():
+        logger.error(f'Error: {args.model} is not a directory')
+        sys.exit(1)
+
+    ftype_map: dict[str, gguf.LlamaFileType] = {
+        "f32": gguf.LlamaFileType.ALL_F32,
+        "f16": gguf.LlamaFileType.MOSTLY_F16,
+        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
+    }
+
+    logger.info(f"Loading model: {dir_model.name}")
+
+    with torch.inference_mode():
+        gemma3_vision_tower = Gemma3VisionTower(
+            dir_model=dir_model,
+            fname_out=args.outfile,
+            ftype=ftype_map[args.outtype],
+            is_big_endian=args.bigendian,
+        )
+        gemma3_vision_tower.write()
+
+
+if __name__ == '__main__':
+    main()
+

examples/llava/minicpmv-cli.cpp

@@ -86,7 +86,11 @@ static struct clip_ctx * clip_init_context(common_params * params) {
     if (prompt.empty()) {
         prompt = "describe the image in detail.";
     }
-    auto * ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
+    struct clip_context_params clip_params = {
+        /* use_gpu */   params->n_gpu_layers != 0,
+        /* verbosity */ params->verbosity,
+    };
+    auto * ctx_clip = clip_init(clip_path, clip_params);
     return ctx_clip;
 }
 

ggml/src/ggml-backend-reg.cpp

@@ -497,7 +497,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
         search_paths.push_back(get_executable_path());
         search_paths.push_back(fs::current_path());
     } else {
-        search_paths.push_back(user_search_path);
+        search_paths.push_back(fs::u8path(user_search_path));
     }
 
     int best_score = 0;
@@ -511,9 +511,9 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
         fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
         for (const auto & entry : dir_it) {
             if (entry.is_regular_file()) {
-                auto filename = entry.path().filename().native();
-                auto ext = entry.path().extension().native();
-                if (filename.find(file_prefix) == 0 && ext == file_extension) {
+                auto filename = entry.path().filename();
+                auto ext = entry.path().extension();
+                if (filename.native().find(file_prefix) == 0 && ext == file_extension) {
                     dl_handle_ptr handle { dl_load_library(entry) };
                     if (!handle && !silent) {
                         GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_str(entry.path()).c_str());
@@ -544,7 +544,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
         // try to load the base backend
         for (const auto & search_path : search_paths) {
             fs::path filename = backend_filename_prefix().native() + name_path.native() + backend_filename_extension().native();
-            fs::path path = search_path.native() + filename.native();
+            fs::path path = search_path / filename;
             if (fs::exists(path)) {
                 return get_reg().load_backend(path, silent);
             }

ggml/src/ggml-cuda/common.cuh

@@ -395,11 +395,11 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
+#if defined(CDNA) || defined(RDNA2) || defined(__gfx906__)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
 #elif defined(RDNA3)
     c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
-#elif defined(__gfx1010__) || defined(__gfx900__)
+#elif defined(RDNA1) || defined(__gfx900__)
     int tmp1;
     int tmp2;
     asm("\n \

ggml/src/ggml-cuda/mmvq.cu

@@ -47,11 +47,89 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
         1;
 }
 
+enum mmvq_parameter_table_id {
+    MMVQ_PARAMETERS_GENERIC = 0,
+    MMVQ_PARAMETERS_GCN,
+    MMVQ_PARAMETERS_RDNA2
+};
+
+static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
+#if defined(RDNA2) || defined(RDNA3)
+    return MMVQ_PARAMETERS_RDNA2;
+#elif defined(GCN) || defined(CDNA)
+    return MMVQ_PARAMETERS_GCN;
+#else
+    return MMVQ_PARAMETERS_GENERIC;
+#endif
+}
+
+static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
+    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3(cc)) {
+        return MMVQ_PARAMETERS_RDNA2;
+    }
+    if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) {
+        return MMVQ_PARAMETERS_GCN;
+    }
+    return MMVQ_PARAMETERS_GENERIC;
+}
+
+static constexpr __host__ __device__ int calc_nwarps(int ncols_y,  mmvq_parameter_table_id table_id) {
+    if (table_id == MMVQ_PARAMETERS_GENERIC) {
+        switch (ncols_y) {
+            case 1:
+            case 2:
+            case 3:
+            case 4:
+                return 4;
+            case 5:
+            case 6:
+            case 7:
+            case 8:
+                return 2;
+            default:
+                return 1;
+        }
+    } else if (table_id == MMVQ_PARAMETERS_GCN) {
+        switch (ncols_y) {
+            case 1:
+            case 2:
+            case 3:
+            case 4:
+                return 2;
+            case 5:
+            case 6:
+            case 7:
+            case 8:
+            default:
+                return 1;
+        }
+    }
+    return 1;
+}
+
+static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int table_id) {
+    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN) {
+        switch (ncols_y) {
+            case 1:
+                return 1;
+            case 2:
+            case 3:
+            case 4:
+            case 5:
+            case 6:
+            case 7:
+            case 8:
+                return 2;
+            default:
+                return 1;
+        }
+    }
+    return 1;
+}
+
 template <ggml_type type, int ncols_y>
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 // tell the compiler to use as many registers as it wants, see nwarps definition below
-__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(calc_nwarps(ncols_y, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
@@ -59,24 +137,20 @@ static __global__ void mul_mat_vec_q(
     constexpr int qk  = ggml_cuda_type_traits<type>::qk;
     constexpr int qi  = ggml_cuda_type_traits<type>::qi;
     constexpr int vdr = get_vdr_mmvq(type);
+    constexpr mmvq_parameter_table_id table_id = get_device_table_id();
+    constexpr int nwarps = calc_nwarps(ncols_y, table_id);
+    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_y, table_id);
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
 
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
-    constexpr int nwarps              = 1;
-    constexpr int rows_per_cuda_block = 1;
-#else
-    constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
-    constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
-
-    const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    const     int tid = warp_size*threadIdx.y + threadIdx.x;
     const     int row0 = rows_per_cuda_block*blockIdx.x;
     const     int blocks_per_row_x = ncols_x / qk;
     const     int blocks_per_col_y = nrows_y / QK8_1;
-    constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi;
+    constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
 
-// partial sum for each thread
+    // partial sum for each thread
     float tmp[ncols_y][rows_per_cuda_block] = {0.0f};
 
     const block_q8_1 * y = (const block_q8_1 *) vy;
@@ -96,7 +170,7 @@ static __global__ void mul_mat_vec_q(
         }
     }
 
-    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE];
+    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][warp_size];
     if (threadIdx.y > 0) {
 #pragma unroll
         for (int j = 0; j < ncols_y; ++j) {
@@ -120,7 +194,7 @@ static __global__ void mul_mat_vec_q(
             for (int l = 0; l < nwarps-1; ++l) {
                 tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
             }
-            tmp[j][i] = warp_reduce_sum(tmp[j][i]);
+            tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
         }
 
         if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < nrows_dst)) {
@@ -129,6 +203,13 @@ static __global__ void mul_mat_vec_q(
     }
 }
 
+static std::pair<dim3, dim3> calc_launch_params(const int ncols_y, const int nrows_x, const int warp_size, const mmvq_parameter_table_id table_id) {
+    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_y, table_id) - 1) / calc_rows_per_block(ncols_y, table_id);
+    const dim3 block_nums(nblocks, 1, 1);
+    const dim3 block_dims(warp_size, calc_nwarps(ncols_y, table_id), 1);
+    return {block_nums, block_dims};
+}
+
 template <ggml_type type>
 static void mul_mat_vec_q_cuda(
     const void * vx, const void * vy, float * dst,
@@ -137,65 +218,67 @@ static void mul_mat_vec_q_cuda(
     GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
     GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
 
-    int id = ggml_cuda_get_device();
-
-    int64_t nwarps = 1;
-    int64_t rows_per_cuda_block = 1;
-
-    if (ggml_cuda_info().devices[id].cc < GGML_CUDA_CC_RDNA2) { // NVIDIA and AMD older than RDNA2
-        switch(ncols_y) {
-            case 1:
-                nwarps = 4;
-                rows_per_cuda_block = 1;
-                break;
-            case 2:
-            case 3:
-            case 4:
-                nwarps = 4;
-                rows_per_cuda_block = 2;
-                break;
-            case 5:
-            case 6:
-            case 7:
-            case 8:
-                nwarps = 2;
-                rows_per_cuda_block = 2;
-                break;
-            default:
-                GGML_ABORT("fatal error");
-                break;
-        }
-    }
-
-    const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block;
-    const dim3 block_nums(nblocks, 1, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+    const int device = ggml_cuda_get_device();
+    const int warp_size = ggml_cuda_info().devices[device].warp_size;
+    const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
 
     switch (ncols_y) {
         case 1:
-            mul_mat_vec_q<type, 1><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+        {
+            constexpr int c_ncols_y = 1;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
+        }
         case 2:
-            mul_mat_vec_q<type, 2><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+        {
+            constexpr int c_ncols_y = 2;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
+        }
         case 3:
-            mul_mat_vec_q<type, 3><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+        {
+            constexpr int c_ncols_y = 3;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
+        }
         case 4:
-            mul_mat_vec_q<type, 4><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+        {
+            constexpr int c_ncols_y = 4;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
+        }
         case 5:
-            mul_mat_vec_q<type, 5><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+        {
+            constexpr int c_ncols_y = 5;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
+        }
         case 6:
-            mul_mat_vec_q<type, 6><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+        {
+            constexpr int c_ncols_y = 6;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
+        }
         case 7:
-            mul_mat_vec_q<type, 7><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+        {
+            constexpr int c_ncols_y = 7;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
+        }
         case 8:
-            mul_mat_vec_q<type, 8><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+        {
+            constexpr int c_ncols_y = 8;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
+        }
         default:
             GGML_ABORT("fatal error");
             break;

ggml/src/ggml-metal/ggml-metal.m

@@ -46,6 +46,7 @@ static struct ggml_backend_device g_ggml_backend_metal_device;
 static struct ggml_backend_metal_device_context {
     id<MTLDevice> mtl_device;
     int           mtl_device_ref_count;
+    id<MTLLibrary> mtl_library;
 
     bool has_simdgroup_reduction;
     bool has_simdgroup_mm;
@@ -57,6 +58,7 @@ static struct ggml_backend_metal_device_context {
 } g_ggml_ctx_dev_main = {
     /*.mtl_device              =*/ nil,
     /*.mtl_device_ref_count    =*/ 0,
+    /*.mtl_library             =*/ nil,
     /*.has_simdgroup_reduction =*/ false,
     /*.has_simdgroup_mm        =*/ false,
     /*.has_residency_sets      =*/ false,
@@ -108,6 +110,11 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     ctx->mtl_device_ref_count--;
 
     if (ctx->mtl_device_ref_count == 0) {
+        if (ctx->mtl_library) {
+            [ctx->mtl_library release];
+            ctx->mtl_library = nil;
+        }
+
         if (ctx->mtl_device) {
             [ctx->mtl_device release];
             ctx->mtl_device = nil;
@@ -495,6 +502,139 @@ static void * ggml_metal_host_malloc(size_t n) {
     return data;
 }
 
+// load library
+//
+// - first check if the library is embedded
+// - then check if the library is in the bundle
+// - if not found, load the source and compile it
+// - if that fails, return NULL
+static id<MTLLibrary> ggml_metal_load_library(id<MTLDevice> device, bool use_bfloat) {
+    id<MTLLibrary> metal_library = nil;
+    NSError * error = nil;
+    NSString * src = nil;
+
+#if GGML_METAL_EMBED_LIBRARY
+    GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+
+    extern const char ggml_metallib_start[];
+    extern const char ggml_metallib_end[];
+
+    src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
+
+#else
+
+#ifdef SWIFT_PACKAGE
+    NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
+#else
+    NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
+
+    NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+    if (path_lib == nil) {
+        // Try to find the resource in the directory where the current binary located.
+        NSString * current_binary = [[NSProcessInfo processInfo] arguments][0];
+        NSString * bin_dir = [current_binary stringByDeletingLastPathComponent];
+        NSString * default_metallib_path = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
+        if ([[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) {
+            GGML_LOG_INFO("%s: found '%s'\n", __func__, [default_metallib_path UTF8String]);
+            NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:default_metallib_path error:&error];
+            if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
+                // Optionally, if this is a symlink, try to resolve it.
+                default_metallib_path = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:default_metallib_path error:&error];
+                if (default_metallib_path && [default_metallib_path length] > 0 && ![[default_metallib_path substringToIndex:1] isEqualToString:@"/"]) {
+                    // It is a relative path, adding the binary directory as directory prefix.
+                    default_metallib_path = [NSString pathWithComponents:@[bin_dir, default_metallib_path]];
+                }
+                if (!default_metallib_path || ![[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) {
+                    // Link to the resource could not be resolved.
+                    default_metallib_path = nil;
+                } else {
+                    GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [default_metallib_path UTF8String]);
+                }
+            }
+        } else {
+            // The resource couldn't be found in the binary's directory.
+            default_metallib_path = nil;
+        }
+        path_lib = default_metallib_path;
+    }
+
+    if (path_lib != nil) {
+        // pre-compiled library found
+        NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+        GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
+        metal_library = [device newLibraryWithURL:libURL error:&error];
+        if (error) {
+            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            return NULL;
+        }
+    } else {
+        GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+
+        NSString * path_source;
+        NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+
+        GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
+
+        if (path_resource) {
+            path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
+        } else {
+            path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+        }
+
+        if (path_source == nil) {
+            GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
+            path_source = @"ggml-metal.metal";
+        }
+
+        GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+        src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
+        if (error) {
+            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            return NULL;
+        }
+    }
+#endif
+
+    if (!metal_library) {
+        @autoreleasepool {
+            // dictionary of preprocessor macros
+            NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+
+            if (use_bfloat) {
+                [prep setObject:@"1" forKey:@"GGML_METAL_USE_BF16"];
+            }
+
+#if GGML_METAL_EMBED_LIBRARY
+            [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
+#endif
+
+            MTLCompileOptions * options = [MTLCompileOptions new];
+            options.preprocessorMacros = prep;
+
+            //[options setFastMathEnabled:false];
+
+            metal_library = [device newLibraryWithSource:src options:options error:&error];
+            if (error) {
+                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+                return NULL;
+            }
+
+#if !__has_feature(objc_arc)
+            [options release];
+#endif
+        }
+    }
+
+#if GGML_METAL_EMBED_LIBRARY
+    [src release];
+#endif // GGML_METAL_EMBED_LIBRARY
+
+    return metal_library;
+}
+
 static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t dev) {
     GGML_LOG_INFO("%s: allocating\n", __func__);
 
@@ -522,136 +662,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
 
     ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
-    id<MTLLibrary> metal_library = nil;
-
     // load library
-    //
-    // - first check if the library is embedded
-    // - then check if the library is in the bundle
-    // - if not found, load the source and compile it
-    // - if that fails, return NULL
-    {
-        NSError * error = nil;
-        NSString * src = nil;
-
-#if GGML_METAL_EMBED_LIBRARY
-        GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
-
-        extern const char ggml_metallib_start[];
-        extern const char ggml_metallib_end[];
-
-        src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-
-#else
-
-#ifdef SWIFT_PACKAGE
-        NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
-#else
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-#endif
-
-        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (path_lib == nil) {
-            // Try to find the resource in the directory where the current binary located.
-            NSString * current_binary = [[NSProcessInfo processInfo] arguments][0];
-            NSString * bin_dir = [current_binary stringByDeletingLastPathComponent];
-            NSString * default_metallib_path = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
-            if ([[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) {
-                GGML_LOG_INFO("%s: found '%s'\n", __func__, [default_metallib_path UTF8String]);
-                NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:default_metallib_path error:&error];
-                if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
-                    // Optionally, if this is a symlink, try to resolve it.
-                    default_metallib_path = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:default_metallib_path error:&error];
-                    if (default_metallib_path && [default_metallib_path length] > 0 && ![[default_metallib_path substringToIndex:1] isEqualToString:@"/"]) {
-                        // It is a relative path, adding the binary directory as directory prefix.
-                        default_metallib_path = [NSString pathWithComponents:@[bin_dir, default_metallib_path]];
-                    }
-                    if (!default_metallib_path || ![[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) {
-                        // Link to the resource could not be resolved.
-                        default_metallib_path = nil;
-                    } else {
-                        GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [default_metallib_path UTF8String]);
-                    }
-                }
-            } else {
-                // The resource couldn't be found in the binary's directory.
-                default_metallib_path = nil;
-            }
-            path_lib = default_metallib_path;
-        }
-
-        if (path_lib != nil) {
-            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
-
-            metal_library = [device newLibraryWithURL:libURL error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return NULL;
-            }
-        } else {
-            GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
-
-            NSString * path_source;
-            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
-
-            GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
-
-            if (path_resource) {
-                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
-            } else {
-                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
-            }
-
-            if (path_source == nil) {
-                GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                path_source = @"ggml-metal.metal";
-            }
-
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
-
-            src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return NULL;
-            }
-        }
-#endif
-
-        if (!metal_library) {
-            @autoreleasepool {
-                // dictionary of preprocessor macros
-                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
-
-                if (ctx_dev->use_bfloat) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_USE_BF16"];
-                }
-
-#if GGML_METAL_EMBED_LIBRARY
-                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
-#endif
-
-                MTLCompileOptions * options = [MTLCompileOptions new];
-                options.preprocessorMacros = prep;
-
-                //[options setFastMathEnabled:false];
-
-                metal_library = [device newLibraryWithSource:src options:options error:&error];
-                if (error) {
-                    GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                    return NULL;
-                }
-
-#if !__has_feature(objc_arc)
-                [options release];
-#endif
-            }
-        }
-
-#if GGML_METAL_EMBED_LIBRARY
-        [src release];
-#endif // GGML_METAL_EMBED_LIBRARY
+    if (ctx_dev->mtl_library == nil) {
+        ctx_dev->mtl_library = ggml_metal_load_library(device, ctx_dev->use_bfloat);
+    }
+    id<MTLLibrary> metal_library = ctx_dev->mtl_library;
+    if (metal_library == nil) {
+        GGML_LOG_ERROR("%s: error: metal library is nil\n", __func__);
+        return NULL;
     }
 
     // print MTL GPU family:
@@ -725,7 +743,6 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
             [metal_function release]; \
             if (error) { \
                 GGML_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
-                [metal_library release]; \
                 return NULL; \
             } \
         } else { \
@@ -1044,8 +1061,6 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,               pool_2d_max_f32,                true);
     }
 
-    [metal_library release];
-
     return ctx;
 }
 

ggml/src/ggml-musa/CMakeLists.txt

@@ -21,7 +21,7 @@ if (MUSAToolkit_FOUND)
     message(STATUS "MUSA Toolkit found")
 
     if (NOT DEFINED MUSA_ARCHITECTURES)
-        set(MUSA_ARCHITECTURES "21;22")
+        set(MUSA_ARCHITECTURES "21;22;31")
     endif()
     message(STATUS "Using MUSA architectures: ${MUSA_ARCHITECTURES}")
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -5,23 +5,24 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-shared FLOAT_TYPE sccache1[BLOCK_SIZE/16][16];
-shared FLOAT_TYPE sccache2[BLOCK_SIZE/16][16];
+shared FLOAT_TYPE sccache1[2][BLOCK_SIZE/16][16];
+shared FLOAT_TYPE sccache2[2][BLOCK_SIZE/16][16];
 
 FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+uint csel = 0;
 
 void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint v_im, const uint ix, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) {
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        csel ^= 1;
 
-        barrier();
         if (!all_threads) { // when we don't have enough blocks to use all threads
             if (i < num_blocks_per_row) {
                 const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]);
-                sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF);
-                sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF);
+                sccache1[csel][ix][itid] = FLOAT_TYPE(scale & 0xF);
+                sccache2[csel][ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF);
             }
             barrier();
 
@@ -29,8 +30,8 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
                 continue;
         } else {
             const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]);
-            sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF);
-            sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF);
+            sccache1[csel][ix][itid] = FLOAT_TYPE(scale & 0xF);
+            sccache2[csel][ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF);
             barrier();
         }
 
@@ -57,22 +58,22 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
             FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
             FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
             [[unroll]] for (int l = 0; l < 2; ++l) {
-                sum1 = fma(FLOAT_TYPE(b0[l]),   sccache1[ix][    8*v_im] * qs_u32_0[l  ],
-                       fma(FLOAT_TYPE(b16[l]),  sccache1[ix][1 + 8*v_im] * qs_u32_0[l+2],
-                       fma(FLOAT_TYPE(b32[l]),  sccache1[ix][2 + 8*v_im] * qs_u32_2[l  ],
-                       fma(FLOAT_TYPE(b48[l]),  sccache1[ix][3 + 8*v_im] * qs_u32_2[l+2],
-                       fma(FLOAT_TYPE(b64[l]),  sccache1[ix][4 + 8*v_im] * qs_u32_4[l  ],
-                       fma(FLOAT_TYPE(b80[l]),  sccache1[ix][5 + 8*v_im] * qs_u32_4[l+2],
-                       fma(FLOAT_TYPE(b96[l]),  sccache1[ix][6 + 8*v_im] * qs_u32_6[l  ],
-                       fma(FLOAT_TYPE(b112[l]), sccache1[ix][7 + 8*v_im] * qs_u32_6[l+2], sum1))))))));
-                sum2 = fma(FLOAT_TYPE(b0[l]),   sccache2[ix][    8*v_im],
-                       fma(FLOAT_TYPE(b16[l]),  sccache2[ix][1 + 8*v_im],
-                       fma(FLOAT_TYPE(b32[l]),  sccache2[ix][2 + 8*v_im],
-                       fma(FLOAT_TYPE(b48[l]),  sccache2[ix][3 + 8*v_im],
-                       fma(FLOAT_TYPE(b64[l]),  sccache2[ix][4 + 8*v_im],
-                       fma(FLOAT_TYPE(b80[l]),  sccache2[ix][5 + 8*v_im],
-                       fma(FLOAT_TYPE(b96[l]),  sccache2[ix][6 + 8*v_im],
-                       fma(FLOAT_TYPE(b112[l]), sccache2[ix][7 + 8*v_im], sum2))))))));
+                sum1 = fma(FLOAT_TYPE(b0[l]),   sccache1[csel][ix][    8*v_im] * qs_u32_0[l  ],
+                       fma(FLOAT_TYPE(b16[l]),  sccache1[csel][ix][1 + 8*v_im] * qs_u32_0[l+2],
+                       fma(FLOAT_TYPE(b32[l]),  sccache1[csel][ix][2 + 8*v_im] * qs_u32_2[l  ],
+                       fma(FLOAT_TYPE(b48[l]),  sccache1[csel][ix][3 + 8*v_im] * qs_u32_2[l+2],
+                       fma(FLOAT_TYPE(b64[l]),  sccache1[csel][ix][4 + 8*v_im] * qs_u32_4[l  ],
+                       fma(FLOAT_TYPE(b80[l]),  sccache1[csel][ix][5 + 8*v_im] * qs_u32_4[l+2],
+                       fma(FLOAT_TYPE(b96[l]),  sccache1[csel][ix][6 + 8*v_im] * qs_u32_6[l  ],
+                       fma(FLOAT_TYPE(b112[l]), sccache1[csel][ix][7 + 8*v_im] * qs_u32_6[l+2], sum1))))))));
+                sum2 = fma(FLOAT_TYPE(b0[l]),   sccache2[csel][ix][    8*v_im],
+                       fma(FLOAT_TYPE(b16[l]),  sccache2[csel][ix][1 + 8*v_im],
+                       fma(FLOAT_TYPE(b32[l]),  sccache2[csel][ix][2 + 8*v_im],
+                       fma(FLOAT_TYPE(b48[l]),  sccache2[csel][ix][3 + 8*v_im],
+                       fma(FLOAT_TYPE(b64[l]),  sccache2[csel][ix][4 + 8*v_im],
+                       fma(FLOAT_TYPE(b80[l]),  sccache2[csel][ix][5 + 8*v_im],
+                       fma(FLOAT_TYPE(b96[l]),  sccache2[csel][ix][6 + 8*v_im],
+                       fma(FLOAT_TYPE(b112[l]), sccache2[csel][ix][7 + 8*v_im], sum2))))))));
             }
             temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
         }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp

@@ -5,20 +5,21 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-shared FLOAT_TYPE sccache[BLOCK_SIZE/16][2][8];
+shared FLOAT_TYPE sccache[2][BLOCK_SIZE/16][2][8];
 
 FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+uint csel = 0;
 
 void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, const uint itid8, const uint v_im, const uint v_im4, const uint v_in, const uint32_t hm_m[4], const uint q_offset, const uint y_offset, const uint s_shift, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) {
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads
-            barrier();
             if (i < num_blocks_per_row)
-                sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32);
+                sccache[csel][ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32);
             barrier();
 
             if (i >= num_blocks_per_row)
@@ -40,8 +41,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co
         const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
 
         if (all_threads) {
-            barrier();
-            sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32);
+            sccache[csel][ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32);
             barrier();
         }
 
@@ -59,14 +59,14 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co
 
             FLOAT_TYPE sum = FLOAT_TYPE(0.0);
             [[unroll]] for (int l = 0; l < 2; ++l) {
-                sum = fma(FLOAT_TYPE(  b0[l]) * sccache[ix][v_im][0], qs_u32_0[l  ] - hmk_0[l  ],
-                      fma(FLOAT_TYPE( b16[l]) * sccache[ix][v_im][1], qs_u32_0[l+2] - hmk_0[l+2],
-                      fma(FLOAT_TYPE( b32[l]) * sccache[ix][v_im][2], qs_u32_2[l  ] - hmk_1[l  ],
-                      fma(FLOAT_TYPE( b48[l]) * sccache[ix][v_im][3], qs_u32_2[l+2] - hmk_1[l+2],
-                      fma(FLOAT_TYPE( b64[l]) * sccache[ix][v_im][4], qs_u32_4[l  ] - hmk_2[l  ],
-                      fma(FLOAT_TYPE( b80[l]) * sccache[ix][v_im][5], qs_u32_4[l+2] - hmk_2[l+2],
-                      fma(FLOAT_TYPE( b96[l]) * sccache[ix][v_im][6], qs_u32_6[l  ] - hmk_3[l  ],
-                      fma(FLOAT_TYPE(b112[l]) * sccache[ix][v_im][7], qs_u32_6[l+2] - hmk_3[l+2], sum))))))));
+                sum = fma(FLOAT_TYPE(  b0[l]) * sccache[csel][ix][v_im][0], qs_u32_0[l  ] - hmk_0[l  ],
+                      fma(FLOAT_TYPE( b16[l]) * sccache[csel][ix][v_im][1], qs_u32_0[l+2] - hmk_0[l+2],
+                      fma(FLOAT_TYPE( b32[l]) * sccache[csel][ix][v_im][2], qs_u32_2[l  ] - hmk_1[l  ],
+                      fma(FLOAT_TYPE( b48[l]) * sccache[csel][ix][v_im][3], qs_u32_2[l+2] - hmk_1[l+2],
+                      fma(FLOAT_TYPE( b64[l]) * sccache[csel][ix][v_im][4], qs_u32_4[l  ] - hmk_2[l  ],
+                      fma(FLOAT_TYPE( b80[l]) * sccache[csel][ix][v_im][5], qs_u32_4[l+2] - hmk_2[l+2],
+                      fma(FLOAT_TYPE( b96[l]) * sccache[csel][ix][v_im][6], qs_u32_6[l  ] - hmk_3[l  ],
+                      fma(FLOAT_TYPE(b112[l]) * sccache[csel][ix][v_im][7], qs_u32_6[l+2] - hmk_3[l+2], sum))))))));
             }
             temp[j][n] = fma(d, sum, temp[j][n]);
         }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp

@@ -6,20 +6,21 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-shared FLOAT_TYPE sccache[BLOCK_SIZE/16][16];
+shared FLOAT_TYPE sccache[2][BLOCK_SIZE/16][16];
 
 FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+uint csel = 0;
 
 void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint ix, const uint ql_offset, const uint qh_offset, const uint s_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) {
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads
-            barrier();
             if (i < num_blocks_per_row)
-                sccache[ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]);
+                sccache[csel][ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]);
             barrier();
 
             if (i >= num_blocks_per_row)
@@ -51,8 +52,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
         const vec4 q3 = vec4(unpack8(q3_u32)) - 32;
 
         if (all_threads) {
-            barrier();
-            sccache[ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]);
+            sccache[csel][ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]);
             barrier();
         }
 
@@ -71,7 +71,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
                 sum[2] = fma(FLOAT_TYPE(by64[l]), q2[l], sum[2]);
                 sum[3] = fma(FLOAT_TYPE(by96[l]), q3[l], sum[3]);
             }
-            temp[j][n] = fma(fma(sum[0], sccache[ix][s_offset], fma(sum[1], sccache[ix][s_offset + 2], fma(sum[2], sccache[ix][s_offset + 4], sum[3] * sccache[ix][s_offset + 6]))), d, temp[j][n]);
+            temp[j][n] = fma(fma(sum[0], sccache[csel][ix][s_offset], fma(sum[1], sccache[csel][ix][s_offset + 2], fma(sum[2], sccache[csel][ix][s_offset + 4], sum[3] * sccache[csel][ix][s_offset + 6]))), d, temp[j][n]);
         }
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -777,7 +777,7 @@ void main() {
         [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
             coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
 
-            [[unroll]] for (uint col = 0; col < BN; col += storestride) {
+            [[unroll]] for (uint col = 0; col < TN; col += storestride) {
                 const uint row_i = dc + cm_col * TN + col + store_c;
                 if (row_i >= _ne1) break;
 

gguf-py/gguf/constants.py

@@ -253,6 +253,7 @@ class MODEL_ARCH(IntEnum):
     MINICPM3         = auto()
     GEMMA            = auto()
     GEMMA2           = auto()
+    GEMMA3           = auto()
     STARCODER2       = auto()
     RWKV6            = auto()
     RWKV6QWEN2       = auto()
@@ -440,6 +441,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.MINICPM3:         "minicpm3",
     MODEL_ARCH.GEMMA:            "gemma",
     MODEL_ARCH.GEMMA2:           "gemma2",
+    MODEL_ARCH.GEMMA3:           "gemma3",
     MODEL_ARCH.STARCODER2:       "starcoder2",
     MODEL_ARCH.RWKV6:            "rwkv6",
     MODEL_ARCH.RWKV6QWEN2:       "rwkv6qwen2",
@@ -1077,6 +1079,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_PRE_NORM,
         MODEL_TENSOR.FFN_POST_NORM,
     ],
+    MODEL_ARCH.GEMMA3: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_PRE_NORM,
+        MODEL_TENSOR.FFN_POST_NORM,
+    ],
     MODEL_ARCH.STARCODER2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

src/llama-arch.cpp

@@ -36,6 +36,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_MINICPM3,         "minicpm3"         },
     { LLM_ARCH_GEMMA,            "gemma"            },
     { LLM_ARCH_GEMMA2,           "gemma2"           },
+    { LLM_ARCH_GEMMA3,           "gemma3"           },
     { LLM_ARCH_STARCODER2,       "starcoder2"       },
     { LLM_ARCH_MAMBA,            "mamba"            },
     { LLM_ARCH_XVERSE,           "xverse"           },
@@ -766,6 +767,26 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
         },
     },
+    {
+        LLM_ARCH_GEMMA3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_STARCODER2,
         {

src/llama-arch.h

@@ -40,6 +40,7 @@ enum llm_arch {
     LLM_ARCH_MINICPM3,
     LLM_ARCH_GEMMA,
     LLM_ARCH_GEMMA2,
+    LLM_ARCH_GEMMA3,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
     LLM_ARCH_XVERSE,

src/llama-model.cpp

@@ -9,6 +9,7 @@
 #include <algorithm>
 #include <cassert>
 #include <cstring>
+#include <cmath>
 #include <functional>
 #include <map>
 #include <sstream>
@@ -864,6 +865,23 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                }
             } break;
+        case LLM_ARCH_GEMMA3:
+            {
+                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 26: type = LLM_TYPE_1B; break;
+                    case 34: type = LLM_TYPE_4B; break;
+                    case 48: type = LLM_TYPE_12B; break;
+                    case 62: type = LLM_TYPE_27B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+
+                hparams.f_attention_scale = type == LLM_TYPE_27B
+                    ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
+                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k));
+            } break;
         case LLM_ARCH_STARCODER2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -2454,6 +2472,35 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
                         layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
 
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+                    }
+                } break;
+            case LLM_ARCH_GEMMA3:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_k}, 0);
+                        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head_k}, 0);
+
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
                         layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
@@ -3650,6 +3697,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: f_clamp_kqv      = %.1e\n",   __func__, hparams.f_clamp_kqv);
         LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n",   __func__, hparams.f_max_alibi_bias);
         LLAMA_LOG_INFO("%s: f_logit_scale    = %.1e\n",   __func__, hparams.f_logit_scale);
+        LLAMA_LOG_INFO("%s: f_attn_scale     = %.1e\n",   __func__, hparams.f_attention_scale);
         LLAMA_LOG_INFO("%s: n_ff             = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_expert         = %u\n",     __func__, hparams.n_expert);
         LLAMA_LOG_INFO("%s: n_expert_used    = %u\n",     __func__, hparams.n_expert_used);
@@ -3923,6 +3971,7 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) {
         case LLM_ARCH_PHIMOE:
         case LLM_ARCH_GEMMA:
         case LLM_ARCH_GEMMA2:
+        case LLM_ARCH_GEMMA3:
         case LLM_ARCH_STARCODER2:
         case LLM_ARCH_OPENELM:
         case LLM_ARCH_GPTNEOX:

src/llama.cpp

@@ -4978,6 +4978,149 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_gemma3() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
+
+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
+
+        // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+        if (ubatch.token) {
+            inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+            cb(inpL, "inp_scaled", -1);
+        }
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        // gemma3 requires different mask for layers using sliding window (SWA)
+        struct ggml_tensor * KQ_mask     = build_inp_KQ_mask(true);
+        struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa(true);
+
+        // "5-to-1 interleaved attention"
+        // 5 layers of local attention followed by 1 layer of global attention
+        static const int sliding_window_pattern = 6;
+
+        for (int il = 0; il < n_layer; ++il) {
+            const bool is_sliding          = (il + 1) % sliding_window_pattern;
+            const float freq_base_l        = is_sliding ? 10000.0f    : freq_base;
+            const float freq_scale_l       = is_sliding ? 1.0f        : freq_scale;
+            struct ggml_tensor * KQ_mask_l = is_sliding ? KQ_mask_swa : KQ_mask;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens);
+                Qcur = llm_build_norm(ctx0, Qcur, hparams,
+                    model.layers[il].attn_q_norm,
+                    NULL,
+                    LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens);
+                Kcur = llm_build_norm(ctx0, Kcur, hparams,
+                    model.layers[il].attn_k_norm,
+                    NULL,
+                    LLM_NORM_RMS, cb, il);
+                cb(Kcur, "Kcur_normed", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask_l, n_tokens, kv_head, n_kv, hparams.f_attention_scale, cb, il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_post_norm", il);
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+            cb(sa_out, "sa_out", il);
+
+            cur = llm_build_norm(ctx0, sa_out, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+            cb(cur, "ffn_post_norm", -1);
+
+            cur = ggml_add(ctx0, cur, sa_out);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
 
     struct ggml_cgraph * build_starcoder2() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
@@ -8298,6 +8441,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_gemma2();
             } break;
+        case LLM_ARCH_GEMMA3:
+            {
+                result = llm.build_gemma3();
+            } break;
         case LLM_ARCH_STARCODER2:
             {
                 result = llm.build_starcoder2();

tests/test-backend-ops.cpp

@@ -4113,7 +4113,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             for (int n_mats : {4, 8}) {
                 for (int n_used : {1, 2, 4}) {
                     for (bool b : {false, true}) {
-                        for (int n : {1, 32}) {
+                        for (int n : {1, 32, 129}) {
                             int m = 512;
                             int k = 256;
                             test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, n_mats, n_used, b, m, n, k));