CUDA: fix crash with partial offloading of MoE (#13439)

* Support InternVL 3 38B and 78B mmproj

* Swap norms in clip.cpp

* Group variables together

.github/workflows/build.yml

@@ -15,7 +15,6 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
   LLAMA_LOG_COLORS: 1
@@ -308,7 +307,7 @@ jobs:
         run: |
           cd build
           # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 2700
+          ctest -L main --verbose --timeout 3600
 
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04

.github/workflows/docker.yml

@@ -42,8 +42,7 @@ jobs:
           - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
           - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
           - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
-          # Note: the intel images are failing due to an out of disk space error
-          # - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
           - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
           # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
           #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }

.github/workflows/release.yml

@@ -16,11 +16,6 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
-# Fine-grant permission
-# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
-permissions:
-  contents: write # for creating release
-
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   CMAKE_ARGS: "-DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=ON -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON"
@@ -416,28 +411,27 @@ jobs:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
-          name: llama-bin-win-cu${{ matrix.cuda }}-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip
+          name: llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
 
       - name: Copy and pack Cuda runtime
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
         run: |
           echo "Cuda install location: ${{ env.CUDA_PATH }}"
           $dst='.\build\bin\cudart\'
           robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\*
+          7z a cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
         uses: actions/upload-artifact@v4
         with:
-          path: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
-          name: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
+          path: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip
 
   windows-sycl:
     runs-on: windows-latest
@@ -646,6 +640,11 @@ jobs:
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
+    # Fine-grant permission
+    # https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+    permissions:
+        contents: write # for creating release
+
     runs-on: ubuntu-latest
 
     needs:

CMakeLists.txt

@@ -252,20 +252,3 @@ configure_file(cmake/llama.pc.in
 
 install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig)
-
-#
-# copy the license files
-#
-
-# Check if running in GitHub Actions
-if(DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true")
-    message(STATUS "Running inside GitHub Actions - copying license files")
-
-    # Copy all files from licenses/ to build/bin/
-    file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*")
-    foreach(LICENSE_FILE ${LICENSE_FILES})
-        get_filename_component(FILENAME ${LICENSE_FILE} NAME)
-        configure_file(${LICENSE_FILE} "${CMAKE_BINARY_DIR}/bin/${FILENAME}" COPYONLY)
-    endforeach()
-endif()
-

README.md

@@ -16,8 +16,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
+- 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
-- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141]((https://github.com/ggml-org/llama.cpp/pull/13141))), `libllava` will be deprecated
+- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim

common/CMakeLists.txt

@@ -119,8 +119,8 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.7.10:
-        GIT_TAG 0309d2a6bf40abda35344a362edc71e06d5009f8
+        # v0.7.19 (+ fancy-regex build fix):
+        GIT_TAG b59f98f85269892a7de3d3641ad155366f13daa6
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE
@@ -144,3 +144,27 @@ endif ()
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 target_link_libraries     (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
+
+
+#
+# copy the license files
+#
+
+# Check if running in GitHub Actions
+if (DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true")
+    message(STATUS "Running inside GitHub Actions - copying license files")
+
+    # Copy all files from licenses/ to build/bin/
+    file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*")
+    foreach(LICENSE_FILE ${LICENSE_FILES})
+        get_filename_component(FILENAME ${LICENSE_FILE} NAME)
+        add_custom_command(
+            POST_BUILD
+            TARGET ${TARGET}
+            COMMAND ${CMAKE_COMMAND} -E copy_if_different
+                "${LICENSE_FILE}"
+                "$<TARGET_FILE_DIR:llama>/${FILENAME}"
+            COMMENT "Copying ${FILENAME} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
+        message(STATUS "Copying ${LICENSE_FILE} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${FILENAME}")
+    endforeach()
+endif()

common/arg.cpp

@@ -40,7 +40,7 @@ using json = nlohmann::ordered_json;
 
 std::initializer_list<enum llama_example> mmproj_examples = {
     LLAMA_EXAMPLE_LLAVA,
-    // TODO: add LLAMA_EXAMPLE_SERVER when it's ready
+    LLAMA_EXAMPLE_SERVER,
 };
 
 static std::string read_file(const std::string & fname) {
@@ -2204,32 +2204,33 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONT_BATCHING"));
     add_opt(common_arg(
         {"--mmproj"}, "FILE",
-        "path to a multimodal projector file. see tools/mtmd/README.md",
+        "path to a multimodal projector file. see tools/mtmd/README.md\n"
+        "note: if -hf is used, this argument can be omitted",
         [](common_params & params, const std::string & value) {
             params.mmproj.path = value;
         }
-    ).set_examples(mmproj_examples));
+    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ"));
     add_opt(common_arg(
         {"--mmproj-url"}, "URL",
         "URL to a multimodal projector file. see tools/mtmd/README.md",
         [](common_params & params, const std::string & value) {
             params.mmproj.url = value;
         }
-    ).set_examples(mmproj_examples));
+    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_URL"));
     add_opt(common_arg(
         {"--no-mmproj"},
         "explicitly disable multimodal projector, useful when using -hf",
         [](common_params & params) {
             params.no_mmproj = true;
         }
-    ).set_examples(mmproj_examples));
+    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ"));
     add_opt(common_arg(
         {"--no-mmproj-offload"},
         "do not offload multimodal projector to GPU",
         [](common_params & params) {
             params.mmproj_use_gpu = false;
         }
-    ).set_examples(mmproj_examples));
+    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ_OFFLOAD"));
     add_opt(common_arg(
         {"--image"}, "FILE",
         "path to an image file. use with multimodal models. Specify multiple times for batching",
@@ -2436,6 +2437,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ));
+    add_opt(common_arg(
+        {"--no-op-offload"},
+        string_format("disable offloading host tensor operations to device (default: %s)", params.no_op_offload ? "true" : "false"),
+        [](common_params & params) {
+            params.no_op_offload = true;
+        }
+    ));
     add_opt(common_arg(
         {"--lora"}, "FNAME",
         "path to LoRA adapter (can be repeated to use multiple adapters)",
@@ -2627,6 +2635,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.i_chunk = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
+    add_opt(common_arg(
+        {"--parse-special"},
+        string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),
+        [](common_params & params) {
+            params.parse_special = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"-pps"},
         string_format("is the prompt shared across parallel sequences (default: %s)", params.is_pp_shared ? "true" : "false"),

common/chat.cpp

@@ -125,7 +125,9 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
             msgs.push_back(msg);
         }
     } catch (const std::exception & e) {
-        throw std::runtime_error("Failed to parse messages: " + std::string(e.what()) + "; messages = " + messages.dump(2));
+        // @ngxson : disable otherwise it's bloating the API response
+        // printf("%s\n", std::string("; messages = ") + messages.dump(2));
+        throw std::runtime_error("Failed to parse messages: " + std::string(e.what()));
     }
 
     return msgs;

common/common.cpp

@@ -1113,6 +1113,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.offload_kqv       = !params.no_kv_offload;
     cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
+    cparams.op_offload        = !params.no_op_offload;
 
     if (params.reranking) {
         cparams.embeddings    = true;

common/common.h

@@ -332,6 +332,7 @@ struct common_params {
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
+    bool no_op_offload     = false; // globally disable offload host tensor operations to device
 
     bool single_turn       = false; // single turn chat conversation
 
@@ -409,6 +410,7 @@ struct common_params {
 
     bool process_output = false; // collect data for the output tensor
     bool compute_ppl    = true;  // whether to compute perplexity
+    bool parse_special  = false; // whether to parse special tokens during imatrix tokenization
 
     // cvector-generator params
     int n_pca_batch = 100;

common/llguidance.cpp

@@ -189,6 +189,7 @@ static LlgTokenizer * llama_sampler_llg_new_tokenizer(const llama_vocab * vocab)
         /* .tokenize_fn                        = */ llama_sampler_llg_tokenize_fn,
         /* .use_approximate_greedy_tokenize_fn = */ false,
         /* .tokenize_user_data                 = */ vocab,
+        /* .slices                             = */ nullptr,
     };
 
     char           error_buffer[1024];

convert_hf_to_gguf.py

@@ -426,7 +426,11 @@ class ModelBase:
             logger.warning(f"Failed to load model config from {dir_model}: {e}")
             logger.warning("Trying to load config.json instead")
             with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-                return json.load(f)
+                config = json.load(f)
+                if "llm_config" in config:
+                    # rename for InternVL
+                    config["text_config"] = config["llm_config"]
+                return config
 
     @classmethod
     def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
@@ -794,6 +798,9 @@ class TextModel(ModelBase):
         if chkhsh == "0e9433cbbb161f89e264eb32e8e64bfe69e834973ffca5d41d3948a604a3e2a3":
             # ref: https://huggingface.co/mistral-community/pixtral-12b
             res = "pixtral"
+        if chkhsh == "d5f1dd6f980fec569fb218a81a7658ac45fc56b38c5a0adeb1c232fbe04ef5ec":
+            # ref: https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base
+            res = "seed-coder"
 
         if res is None:
             logger.warning("\n")
@@ -1388,10 +1395,10 @@ class BaichuanModel(TextModel):
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
         self.gguf_writer.add_file_type(self.ftype)
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         head_count = self.hparams["num_attention_heads"]
@@ -1512,10 +1519,10 @@ class XverseModel(TextModel):
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
         self.gguf_writer.add_file_type(self.ftype)
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -1828,10 +1835,10 @@ class LlamaModel(TextModel):
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     @staticmethod
     def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
@@ -2206,10 +2213,10 @@ class DeciModel(TextModel):
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     @staticmethod
     def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
@@ -2449,10 +2456,10 @@ class MiniCPMModel(TextModel):
         logit_scale = self.hparams["hidden_size"] / self.hparams["dim_model_base"]
         self.gguf_writer.add_logit_scale(logit_scale)
         logger.info(f"gguf: (minicpm) logit_scale = {logit_scale}")
-        if self.hparams.get("rope_scaling") is not None:
-            if self.hparams["rope_scaling"].get("type") == "longrope":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE)
-                logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}")
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "longrope":
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE)
+            logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}")
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         rope_dims = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
@@ -2597,15 +2604,20 @@ class Qwen2Model(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self._try_set_pooling_type()
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "yarn":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if self.hf_arch == "Qwen2Model":
             name = f"model.{name}"  # map to Qwen2ForCausalLM tensors
+        if "language_model." in name:
+            name = name.replace("language_model.", "") # for InternVL
+        if name.startswith("mlp") or name.startswith("vision_model"):
+            # skip visual tensors
+            return []
         yield from super().modify_tensors(data_torch, name, bid)
 
 
@@ -2709,6 +2721,62 @@ class Qwen2VLVisionModel(VisionModel):
         return [] # skip other tensors
 
 
+@ModelBase.register("InternVisionModel")
+class InternVisionModel(VisionModel):
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.INTERNVL)
+        self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
+        # hidden_act
+        if hparams["hidden_act"] == "silu":
+            self.gguf_writer.add_vision_use_silu(True)
+        elif hparams["hidden_act"] == "gelu":
+            self.gguf_writer.add_vision_use_gelu(True)
+        else:
+            raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")
+        # downsample_ratio
+        downsample_ratio = self.global_config.get("downsample_ratio")
+        assert downsample_ratio is not None
+        self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del bid, name, n_dims  # unused
+        if ".patch_embd." in new_name:
+            return gguf.GGMLQuantizationType.F16
+        if ".position_embd." in new_name:
+            return gguf.GGMLQuantizationType.F32
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        if name.startswith("vision_model") or name.startswith("mlp"):
+            # process visual tensors
+            # correct name
+            if name.startswith("vision_model"):
+                name = "vision_tower." + name
+            if (".ls" in name or "position_embedding" in name) and not name.endswith(".weight"):
+                name += ".weight"
+            # split QKV tensors if needed
+            if ".qkv." in name:
+                if data_torch.ndim == 2: # weight
+                    c3, _ = data_torch.shape
+                else: # bias
+                    c3 = data_torch.shape[0]
+                assert c3 % 3 == 0
+                c = c3 // 3
+                wq = data_torch[:c]
+                wk = data_torch[c: c * 2]
+                wv = data_torch[c * 2:]
+                return [
+                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.q_proj")), wq),
+                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.k_proj")), wk),
+                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.v_proj")), wv),
+                ]
+            return [(self.map_tensor_name(name), data_torch)]
+        return [] # skip other tensors
+
+
 @ModelBase.register("WavTokenizerDec")
 class WavTokenizerDecModel(TextModel):
     model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC
@@ -2763,11 +2831,11 @@ class Qwen2MoeModel(TextModel):
             logger.info(f"gguf: expert shared feed forward length = {shared_expert_intermediate_size}")
         # YaRN is not enabled by default
         # To enable it, please refer to this guide: https://huggingface.co/Qwen/Qwen3-30B-A3B#processing-long-texts
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "yarn":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
 
     _experts: list[dict[str, Tensor]] | None = None
 
@@ -3035,7 +3103,7 @@ class Phi3MiniModel(TextModel):
 
         scale = max_pos_embds / orig_max_pos_embds
 
-        rope_scaling_type = rope_scaling.get('type', '').lower()
+        rope_scaling_type = rope_scaling.get('rope_type', rope_scaling.get('type', '')).lower()
         if len(rope_scaling_type) == 0:
             raise KeyError('Missing the required key rope_scaling.type')
 
@@ -3347,10 +3415,10 @@ class InternLM2Model(TextModel):
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
         self.gguf_writer.add_file_type(self.ftype)
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         num_heads = self.hparams["num_attention_heads"]
@@ -3360,6 +3428,11 @@ class InternLM2Model(TextModel):
         head_dim = n_embd // num_heads
         num_groups = num_heads // q_per_kv
 
+        name = name.replace("language_model.", "") # InternVL
+        if name.startswith("mlp") or name.startswith("vision_model"):
+            # skip visual tensors
+            return []
+
         if bid is not None and f"model.layers.{bid}.attention.wqkv" in name:
             qkv = data_torch
 
@@ -3425,14 +3498,18 @@ class InternLM3Model(TextModel):
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "linear" or self.hparams["rope_scaling"].get("rope_type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
+        name = name.replace("language_model.", "") # InternVL
+        if name.startswith("mlp") or name.startswith("vision_model"):
+            # skip visual tensors
+            return []
         if name.endswith(("q_proj.weight", "q_proj.bias")):
             data_torch = LlamaModel.permute(data_torch, n_head, n_head)
         if name.endswith(("k_proj.weight", "k_proj.bias")):
@@ -4866,12 +4943,12 @@ class DeepseekV2Model(TextModel):
 
         self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
 
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "yarn":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
-                self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * hparams["rope_scaling"]["mscale_all_dim"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+            self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * rope_scaling["mscale_all_dim"])
 
     _experts: list[dict[str, Tensor]] | None = None
 
@@ -5363,11 +5440,11 @@ class Glm4Model(TextModel):
         super().set_gguf_parameters()
         rope_dim = self.hparams["head_dim"]
         self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
-        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
-            if self.hparams["rope_scaling"].get("type") == "yarn":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
 
 
 @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
@@ -5600,10 +5677,10 @@ class ExaoneModel(TextModel):
         rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"], optional=True)
         rotary_factor = rotary_factor if rotary_factor is not None else 1.0
         self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
-        if hparams.get("rope_scaling") is not None and "factor" in hparams["rope_scaling"]:
-            if hparams["rope_scaling"].get("type") == "linear":
-                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-                self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
@@ -5706,10 +5783,11 @@ class BailingMoeModel(TextModel):
         rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"]
 
         self.gguf_writer.add_rope_dimension_count(rope_dim)
-        if (self.hparams.get("rope_scaling") or {}).get("type") == "yarn" and "factor" in self.hparams["rope_scaling"]:
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
             self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-            self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
-            self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
         else:
             self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
         self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])

convert_hf_to_gguf_update.py

@@ -116,6 +116,7 @@ models = [
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
     {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
     {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
+    {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
 ]
 
 

docs/multimodal.md

@@ -0,0 +1,77 @@
+# Multimodal
+
+llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools support this feature:
+- [llama-mtmd-cli](../tools/mtmd/README.md)
+- [llama-server](../tools/server/README.md) via OpenAI-compatible `/chat/completions` API
+
+To enable it, can use use one of the 2 methods below:
+
+- Use `-hf` option with a supported model (see a list of pre-quantized model below)
+    - To load a model using `-hf` while disabling multimodal, use `--no-mmproj`
+    - To load a model using `-hf` while using a custom mmproj file, use `--mmproj local_file.gguf`
+- Use `-m model.gguf` option with `--mmproj file.gguf` to specify text and multimodal projector respectively
+
+By default, multimodal projector will be offloaded to GPU. To disable this, add `--no-mmproj-offload`
+
+For example:
+
+```sh
+# simple usage with CLI
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+
+# simple usage with server
+llama-server -hf ggml-org/gemma-3-4b-it-GGUF
+
+# using local file
+llama-server -m gemma-3-4b-it-Q4_K_M.gguf --mmproj mmproj-gemma-3-4b-it-Q4_K_M.gguf
+
+# no GPU offload
+llama-server -hf ggml-org/gemma-3-4b-it-GGUF --no-mmproj-offload
+```
+
+## Pre-quantized models
+
+These are ready-to-use models, most of them come with `Q4_K_M` quantization by default.
+
+Replaces the `(tool_name)` with the name of binary you want to use. For example, `llama-mtmd-cli` or `llama-server`
+
+NOTE: some models may require large context window, for example: `-c 8192`
+
+```sh
+# Gemma 3
+(tool_name) -hf ggml-org/gemma-3-4b-it-GGUF
+(tool_name) -hf ggml-org/gemma-3-12b-it-GGUF
+(tool_name) -hf ggml-org/gemma-3-27b-it-GGUF
+
+# SmolVLM
+(tool_name) -hf ggml-org/SmolVLM-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM-256M-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM-500M-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
+
+# Pixtral 12B
+(tool_name) -hf ggml-org/pixtral-12b-GGUF
+
+# Qwen 2 VL
+(tool_name) -hf ggml-org/Qwen2-VL-2B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2-VL-7B-Instruct-GGUF
+
+# Qwen 2.5 VL
+(tool_name) -hf ggml-org/Qwen2.5-VL-3B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2.5-VL-7B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2.5-VL-32B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2.5-VL-72B-Instruct-GGUF
+
+# Mistral Small 3.1 24B (IQ2_M quantization)
+(tool_name) -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF
+
+# InternVL 2.5 and 3
+(tool_name) -hf ggml-org/InternVL2_5-1B-GGUF
+(tool_name) -hf ggml-org/InternVL2_5-4B-GGUF
+(tool_name) -hf ggml-org/InternVL3-1B-Instruct-GGUF
+(tool_name) -hf ggml-org/InternVL3-2B-Instruct-GGUF
+(tool_name) -hf ggml-org/InternVL3-8B-Instruct-GGUF
+(tool_name) -hf ggml-org/InternVL3-14B-Instruct-GGUF
+```

ggml/include/ggml-backend.h

@@ -248,7 +248,7 @@ extern "C" {
         // preferrably to run on the same backend as the buffer
         ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false, true);
 
         // initialize buffers from a max size graph (optional)
         reserve_graph = build_graph(sched, max_batch_size);
@@ -289,7 +289,7 @@ extern "C" {
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
     // Initialize a backend scheduler, backends with low index are given priority over backends with high index
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
     // Initialize backend buffers from a measure graph

ggml/src/ggml-backend.cpp

@@ -674,6 +674,8 @@ struct ggml_backend_sched {
     char * context_buffer;
     size_t context_buffer_size;
 
+    bool op_offload;
+
     int debug;
 };
 
@@ -766,7 +768,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
-            if (src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
+            if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
                 for (int b = 0; b < src_backend_id; b++) {
                     if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                         SET_CAUSE(tensor, "1.off");
@@ -1452,7 +1454,8 @@ ggml_backend_sched_t ggml_backend_sched_new(
         ggml_backend_buffer_type_t * bufts,
         int n_backends,
         size_t graph_size,
-        bool parallel) {
+        bool parallel,
+        bool op_offload) {
     GGML_ASSERT(n_backends > 0);
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
     GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
@@ -1497,6 +1500,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
     }
 
     sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
+    sched->op_offload = op_offload;
 
     ggml_backend_sched_reset(sched);
 

ggml/src/ggml-cuda/CMakeLists.txt

@@ -118,7 +118,7 @@ if (CUDAToolkit_FOUND)
 
     set(CUDA_CXX_FLAGS "")
 
-    set(CUDA_FLAGS -use_fast_math)
+    set(CUDA_FLAGS -use_fast_math -extended-lambda)
 
     if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
         # Options are:

ggml/src/ggml-cuda/common.cuh

@@ -296,6 +296,25 @@ static __device__ void no_device_code(
 #define NO_DEVICE_CODE //GGML_ABORT("NO_DEVICE_CODE not valid in host code.")
 #endif // __CUDA_ARCH__
 
+// The compiler is always able to unroll loops if they contain continue expressions.
+// In such cases loop unrolling can still be achieved via recursion:
+template <int n>
+struct ggml_cuda_unroll {
+    template <typename Func, typename... Args>
+    __device__ void operator()(const Func & f, Args... args) const {
+        f(n - 1, args...);
+        ggml_cuda_unroll<n - 1>{}(f, args...);
+    }
+};
+
+template <>
+struct ggml_cuda_unroll<1> {
+    template <typename Func, typename... Args>
+    __device__ void operator()(const Func & f, Args... args) const {
+        f(0, args...);
+    }
+};
+
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_sum(int x) {
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE

ggml/src/ggml-cuda/cp-async.cuh

@@ -2,6 +2,17 @@
 
 #include "common.cuh"
 
+
+static __device__ __forceinline__ unsigned int ggml_cuda_cvta_generic_to_shared(void * generic_ptr) {
+#ifdef CP_ASYNC_AVAILABLE
+    return __cvta_generic_to_shared(generic_ptr);
+#else
+    GGML_UNUSED(generic_ptr);
+    NO_DEVICE_CODE;
+    return 0;
+#endif // CP_ASYNC_AVAILABLE
+}
+
 // Copies data from global to shared memory, cg == cache global.
 // Both the src and dst pointers must be aligned to 16 bit.
 // Shared memory uses 32 bit addressing, the pointer is passed as unsigned int.

ggml/src/ggml-cuda/fattn-common.cuh

@@ -516,7 +516,7 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
         nullptr;
 }
 
-template<int D, int ncols1, int ncols2, int KQ_stride> // D == head size
+template<int D, int ncols1, int ncols2> // D == head size
 __launch_bounds__(D, 1)
 static __global__ void flash_attn_stream_k_fixup(
         float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne11) {
@@ -665,13 +665,13 @@ static void on_no_fattn_vec_case(const int D) {
         fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for all combinations of q4_0, q4_1, q5_0, q5_1, q8_0, and f16.\n");
         GGML_ABORT("fatal error");
     } else {
-        fprintf(stderr, "Unsupported KV type combination for head_size 256.\n");
+        fprintf(stderr, "Unsupported KV type combination for head_size %d.\n", D);
         fprintf(stderr, "Only f16 is supported.\n");
         GGML_ABORT("fatal error");
     }
 }
 
-template <int D, int ncols1, int ncols2, int KQ_stride>
+template <int DV, int ncols1, int ncols2>
 void launch_fattn(
     ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared,
     const int KQ_row_granularity, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE
@@ -691,7 +691,7 @@ void launch_fattn(
 
     GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
     GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
-                                "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
+        "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
 
     GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
 
@@ -754,10 +754,13 @@ void launch_fattn(
     const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3];
 
     const dim3 block_dim(warp_size, nwarps, 1);
+    int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
+    CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
+
     dim3 blocks_num;
     if (stream_k) {
         // For short contexts it can be faster to have the SMs work on whole tiles because this lets us skip the fixup.
-        const int max_blocks = 2*nsm;
+        const int max_blocks = max_blocks_per_sm*nsm;
         const int tiles_nwaves = (ntiles_total + max_blocks - 1) / max_blocks;
         const int tiles_efficiency_percent = 100 * ntiles_total / (max_blocks*tiles_nwaves);
 
@@ -769,14 +772,11 @@ void launch_fattn(
         blocks_num.y = 1;
         blocks_num.z = 1;
 
-        dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + D) * sizeof(float));
+        dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + DV) * sizeof(float));
     } else {
         GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0);
         const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
 
-        int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
-        CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
-
         // parallel_blocks should be at least large enough to achieve max. occupancy for a single wave:
         parallel_blocks = std::max((nsm * max_blocks_per_sm) / ntiles_total, 1);
 
@@ -853,19 +853,19 @@ void launch_fattn(
 
     if (stream_k) {
         if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
-            const dim3 block_dim_combine(D, 1, 1);
+            const dim3 block_dim_combine(DV, 1, 1);
             const dim3 blocks_num_combine = {blocks_num.x, ncols1, ncols2};
 
-            flash_attn_stream_k_fixup<D, ncols1, ncols2, KQ_stride>
+            flash_attn_stream_k_fixup<DV, ncols1, ncols2>
                 <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
                 ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], K->ne[1]);
         }
     } else if (parallel_blocks > 1) {
-        const dim3 block_dim_combine(D, 1, 1);
+        const dim3 block_dim_combine(DV, 1, 1);
         const dim3 blocks_num_combine(Q->ne[1], 1, blocks_num.z);
         const size_t nbytes_shared_combine = parallel_blocks*sizeof(float2);
 
-        flash_attn_combine_results<D>
+        flash_attn_combine_results<DV>
             <<<blocks_num_combine, block_dim_combine, nbytes_shared_combine, main_stream>>>
             (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data, parallel_blocks);
     }

ggml/src/ggml-cuda/fattn-tile-f16.cu

@@ -307,7 +307,7 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
             fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, -1>
+            launch_fattn<D, cols_per_block, 1>
                 (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
         } break;
         case 128: {
@@ -315,7 +315,7 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
             fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, -1>
+            launch_fattn<D, cols_per_block, 1>
                 (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
         } break;
         default: {

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -318,7 +318,7 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
             fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, -1>
+            launch_fattn<D, cols_per_block, 1>
                 (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
         } break;
         case 128: {
@@ -326,7 +326,7 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
             fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, -1>
+            launch_fattn<D, cols_per_block, 1>
                 (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
         } break;
         default: {

ggml/src/ggml-cuda/fattn-vec-f16.cuh

@@ -168,6 +168,7 @@ static __global__ void flash_attn_vec_ext_f16(
     for (int j = 0; j < ncols; ++j) {
         KQ[j*D + tid] = -HALF_MAX_HALF;
     }
+    __syncthreads();
 
     half2 VKQ[ncols] = {{0.0f, 0.0f}};
 
@@ -315,7 +316,7 @@ void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx,
     constexpr bool need_f16_K = D != 128;
     constexpr bool need_f16_V = D != 128 && D != 64;
     constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
+    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }
 
 template <int D, ggml_type type_K, ggml_type type_V>

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -310,7 +310,7 @@ void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx,
     constexpr bool need_f16_K = D != 128;
     constexpr bool need_f16_V = D != 128 && D != 64;
     constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
+    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }
 
 template <int D, ggml_type type_K, ggml_type type_V>

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -490,7 +490,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm
         fattn_kernel = flash_attn_ext_f16<
             D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>;
     }
-    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size);
+    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size);
 }
 
 void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/fattn.cu

@@ -8,58 +8,32 @@
 #include "fattn-wmma-f16.cuh"
 #include "fattn.cuh"
 
-template <int D, int ncols2>
+template <int DKQ, int DV, int ncols2>
 static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * Q = dst->src[0];
 
-    if (Q->ne[1] <= 8/ncols2) {
-        ggml_cuda_flash_attn_ext_mma_f16_case<D, 8/ncols2, ncols2>(ctx, dst);
-        return;
+    if constexpr (ncols2 <= 8) {
+        if (Q->ne[1] <= 8/ncols2) {
+            ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 8/ncols2, ncols2>(ctx, dst);
+            return;
+        }
     }
 
     if (Q->ne[1] <= 16/ncols2) {
-        ggml_cuda_flash_attn_ext_mma_f16_case<D, 16/ncols2, ncols2>(ctx, dst);
+        ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
         return;
     }
 
     if (Q->ne[1] <= 32/ncols2) {
-        ggml_cuda_flash_attn_ext_mma_f16_case<D, 32/ncols2, ncols2>(ctx, dst);
+        ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 32/ncols2, ncols2>(ctx, dst);
         return;
     }
 
-    ggml_cuda_flash_attn_ext_mma_f16_case<D, 64/ncols2, ncols2>(ctx, dst);
+    ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 64/ncols2, ncols2>(ctx, dst);
 }
 
-template <int ncols2>
-static void ggml_cuda_flash_attn_ext_mma_f16_switch_hs(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-
-    switch (Q->ne[0]) {
-        case 64:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 64, ncols2>(ctx, dst);
-            break;
-        case 80:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 80, ncols2>(ctx, dst);
-            break;
-        case 96:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 96, ncols2>(ctx, dst);
-            break;
-        case 112:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<112, ncols2>(ctx, dst);
-            break;
-        case 128:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<128, ncols2>(ctx, dst);
-            break;
-        case 256:
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<256, ncols2>(ctx, dst);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-            break;
-    }
-}
-
-static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+template <int DKQ, int DV>
+static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * KQV  = dst;
     const ggml_tensor * Q    = dst->src[0];
     const ggml_tensor * K    = dst->src[1];
@@ -68,27 +42,79 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
     float max_bias = 0.0f;
     memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
 
-    const float use_gqa_opt = mask && max_bias == 0.0f;
+    const bool use_gqa_opt = mask && max_bias == 0.0f;
 
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
     const int gqa_ratio = Q->ne[2] / K->ne[2];
 
     if (use_gqa_opt && gqa_ratio % 8 == 0) {
-        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<8>(ctx, dst);
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 8>(ctx, dst);
         return;
     }
 
-    if (use_gqa_opt && gqa_ratio == 4) {
-        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<4>(ctx, dst);
+    if (use_gqa_opt && gqa_ratio % 4 == 0) {
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 4>(ctx, dst);
         return;
     }
 
-    if (use_gqa_opt && gqa_ratio == 2) {
-        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<2>(ctx, dst);
+    if (use_gqa_opt && gqa_ratio % 2 == 0) {
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
         return;
     }
 
-    ggml_cuda_flash_attn_ext_mma_f16_switch_hs<1>(ctx, dst);
+    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 1>(ctx, dst);
+}
+
+static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV  = dst;
+    const ggml_tensor * Q    = dst->src[0];
+    const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * V    = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];
+
+    switch (Q->ne[0]) {
+        case 64:
+            GGML_ASSERT(V->ne[0] == 64);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 64,  64>(ctx, dst);
+            break;
+        case 80:
+            GGML_ASSERT(V->ne[0] == 80);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 80,  80>(ctx, dst);
+            break;
+        case 96:
+            GGML_ASSERT(V->ne[0] == 96);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 96,  96>(ctx, dst);
+            break;
+        case 112:
+            GGML_ASSERT(V->ne[0] == 112);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<112, 112>(ctx, dst);
+            break;
+        case 128:
+            GGML_ASSERT(V->ne[0] == 128);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<128, 128>(ctx, dst);
+            break;
+        case 256:
+            GGML_ASSERT(V->ne[0] == 256);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
+            break;
+        case 576: {
+            // For Deepseek, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
+            GGML_ASSERT(V->ne[0] == 512);
+            float max_bias = 0.0f;
+            memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+
+            const bool use_gqa_opt = mask && max_bias == 0.0f;
+            GGML_ASSERT(use_gqa_opt);
+
+            GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
+            const int gqa_ratio = Q->ne[2] / K->ne[2];
+            GGML_ASSERT(gqa_ratio % 16 == 0);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+        } break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
 }
 
 #define FATTN_VEC_F16_CASE(D, type_K, type_V)                               \
@@ -299,7 +325,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
     const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
     const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion;
-    const bool can_use_vector_kernel = Q->ne[0] % (2*warp_size) == 0;
+    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
     if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
         if (prec == GGML_PREC_DEFAULT) {
             ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);

ggml/src/ggml-cuda/getrows.cu

@@ -10,10 +10,11 @@ static __global__ void k_get_rows(
         /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
-    const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2;
-    const int i10 =  blockDim.y*blockIdx.y + threadIdx.y;
-    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
-    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+    // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
+    const int i00 = (blockIdx.y * blockDim.x + threadIdx.x)*2;
+    const int i10 =  blockIdx.x;
+    const int i11 =  blockIdx.z / ne12;
+    const int i12 =  blockIdx.z % ne12;
 
     if (i00 >= ne00) {
         return;
@@ -46,10 +47,11 @@ static __global__ void k_get_rows_float(
         /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
-    const int i00 =  blockIdx.x*blockDim.x + threadIdx.x;
-    const int i10 =  blockDim.y*blockIdx.y + threadIdx.y;
-    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
-    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+    // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
+    const int i00 = blockIdx.y * blockDim.x + threadIdx.x;
+    const int i10 = blockIdx.x;
+    const int i11 = blockIdx.z / ne12;
+    const int i12 = blockIdx.z % ne12;
 
     if (i00 >= ne00) {
         return;
@@ -94,8 +96,8 @@ static void get_rows_cuda_q(
         const size_t nb1, const size_t nb2, const size_t nb3,
         cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
-    const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
-    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+    const int block_num_y = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
+    const dim3 block_nums(ne10, block_num_y, ne11*ne12);
 
     // strides in elements
     // const size_t s0 = nb0 / sizeof(dst_t);
@@ -127,8 +129,8 @@ static void get_rows_cuda_float(
         const size_t nb1, const size_t nb2, const size_t nb3,
         cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
-    const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
-    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+    const int block_num_y = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
+    const dim3 block_nums(ne10, block_num_y, ne11*ne12);
 
     // strides in elements
     // const size_t s0 = nb0 / sizeof(dst_t);

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -1909,13 +1909,19 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
+    // If src0 is a temporary compute buffer it may have some padding that needs to be cleared for mul_mat_vec_q or mul_mat_q.
+    // But if src0 is also a view of another tensor then this cannot be done safely because it may overwrite valid tensor data.
+    // Therefore, in such cases use cuBLAS.
+    const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE
+        && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src;
+
     bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
-    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
-    bool use_mul_mat_q     = ggml_is_quantized(src0->type)
+    bool use_mul_mat_q     = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
     bool any_gpus_with_slow_fp16   = false;
@@ -3215,16 +3221,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             return false;
 #endif // FLASH_ATTN_AVAILABLE
             if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
-                // different head sizes of K and V are not supported yet
-                return false;
+                const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
+                if (!new_mma_available(cc) || cc < GGML_CUDA_CC_AMPERE) {
+                    return false;
+                }
+                const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2];
+                return op->src[1]->ne[0] == 576 && op->src[2]->ne[0] == 512 && op->src[3] && gqa_ratio % 16 == 0;
             }
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
-            if (op->src[0]->ne[0] == 576) {
-                // DeepSeek MLA
-                return false;
-            }
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

ggml/src/ggml-cuda/mmq.cu

@@ -91,11 +91,11 @@ void ggml_cuda_mul_mat_q(
 
     // If src0 is a temporary compute buffer, clear any potential padding.
     if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
-        GGML_ASSERT(ggml_is_contiguously_allocated(src0));
-        GGML_ASSERT(!src0->view_src);
         const size_t size_data  = ggml_nbytes(src0);
         const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
         if (size_alloc > size_data) {
+            GGML_ASSERT(ggml_is_contiguously_allocated(src0));
+            GGML_ASSERT(!src0->view_src);
             CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
         }
     }

ggml/src/ggml-cuda/mmvq.cu

@@ -515,11 +515,11 @@ void ggml_cuda_mul_mat_vec_q(
 
     // If src0 is a temporary compute buffer, clear any potential padding.
     if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
-        GGML_ASSERT(ggml_is_contiguously_allocated(src0));
-        GGML_ASSERT(!src0->view_src);
         const size_t size_data  = ggml_nbytes(src0);
         const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
         if (size_alloc > size_data) {
+            GGML_ASSERT(ggml_is_contiguously_allocated(src0));
+            GGML_ASSERT(!src0->view_src);
             CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
         }
     }

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 1, 8);
-DECL_FATTN_MMA_F16_CASE(80, 1, 8);
-DECL_FATTN_MMA_F16_CASE(96, 1, 8);
-DECL_FATTN_MMA_F16_CASE(112, 1, 8);
-DECL_FATTN_MMA_F16_CASE(128, 1, 8);
-DECL_FATTN_MMA_F16_CASE(256, 1, 8);
+DECL_FATTN_MMA_F16_CASE(64, 64, 1, 8);
+DECL_FATTN_MMA_F16_CASE(80, 80, 1, 8);
+DECL_FATTN_MMA_F16_CASE(96, 96, 1, 8);
+DECL_FATTN_MMA_F16_CASE(112, 112, 1, 8);
+DECL_FATTN_MMA_F16_CASE(128, 128, 1, 8);
+DECL_FATTN_MMA_F16_CASE(256, 256, 1, 8);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 16, 1);
-DECL_FATTN_MMA_F16_CASE(80, 16, 1);
-DECL_FATTN_MMA_F16_CASE(96, 16, 1);
-DECL_FATTN_MMA_F16_CASE(112, 16, 1);
-DECL_FATTN_MMA_F16_CASE(128, 16, 1);
-DECL_FATTN_MMA_F16_CASE(256, 16, 1);
+DECL_FATTN_MMA_F16_CASE(64, 64, 16, 1);
+DECL_FATTN_MMA_F16_CASE(80, 80, 16, 1);
+DECL_FATTN_MMA_F16_CASE(96, 96, 16, 1);
+DECL_FATTN_MMA_F16_CASE(112, 112, 16, 1);
+DECL_FATTN_MMA_F16_CASE(128, 128, 16, 1);
+DECL_FATTN_MMA_F16_CASE(256, 256, 16, 1);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 16, 2);
-DECL_FATTN_MMA_F16_CASE(80, 16, 2);
-DECL_FATTN_MMA_F16_CASE(96, 16, 2);
-DECL_FATTN_MMA_F16_CASE(112, 16, 2);
-DECL_FATTN_MMA_F16_CASE(128, 16, 2);
-DECL_FATTN_MMA_F16_CASE(256, 16, 2);
+DECL_FATTN_MMA_F16_CASE(64, 64, 16, 2);
+DECL_FATTN_MMA_F16_CASE(80, 80, 16, 2);
+DECL_FATTN_MMA_F16_CASE(96, 96, 16, 2);
+DECL_FATTN_MMA_F16_CASE(112, 112, 16, 2);
+DECL_FATTN_MMA_F16_CASE(128, 128, 16, 2);
+DECL_FATTN_MMA_F16_CASE(256, 256, 16, 2);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 16, 4);
-DECL_FATTN_MMA_F16_CASE(80, 16, 4);
-DECL_FATTN_MMA_F16_CASE(96, 16, 4);
-DECL_FATTN_MMA_F16_CASE(112, 16, 4);
-DECL_FATTN_MMA_F16_CASE(128, 16, 4);
-DECL_FATTN_MMA_F16_CASE(256, 16, 4);
+DECL_FATTN_MMA_F16_CASE(64, 64, 16, 4);
+DECL_FATTN_MMA_F16_CASE(80, 80, 16, 4);
+DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
+DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
+DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
+DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 2, 4);
-DECL_FATTN_MMA_F16_CASE(80, 2, 4);
-DECL_FATTN_MMA_F16_CASE(96, 2, 4);
-DECL_FATTN_MMA_F16_CASE(112, 2, 4);
-DECL_FATTN_MMA_F16_CASE(128, 2, 4);
-DECL_FATTN_MMA_F16_CASE(256, 2, 4);
+DECL_FATTN_MMA_F16_CASE(64, 64, 2, 4);
+DECL_FATTN_MMA_F16_CASE(80, 80, 2, 4);
+DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
+DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
+DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
+DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 2, 8);
-DECL_FATTN_MMA_F16_CASE(80, 2, 8);
-DECL_FATTN_MMA_F16_CASE(96, 2, 8);
-DECL_FATTN_MMA_F16_CASE(112, 2, 8);
-DECL_FATTN_MMA_F16_CASE(128, 2, 8);
-DECL_FATTN_MMA_F16_CASE(256, 2, 8);
+DECL_FATTN_MMA_F16_CASE(64, 64, 2, 8);
+DECL_FATTN_MMA_F16_CASE(80, 80, 2, 8);
+DECL_FATTN_MMA_F16_CASE(96, 96, 2, 8);
+DECL_FATTN_MMA_F16_CASE(112, 112, 2, 8);
+DECL_FATTN_MMA_F16_CASE(128, 128, 2, 8);
+DECL_FATTN_MMA_F16_CASE(256, 256, 2, 8);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 32, 1);
-DECL_FATTN_MMA_F16_CASE(80, 32, 1);
-DECL_FATTN_MMA_F16_CASE(96, 32, 1);
-DECL_FATTN_MMA_F16_CASE(112, 32, 1);
-DECL_FATTN_MMA_F16_CASE(128, 32, 1);
-DECL_FATTN_MMA_F16_CASE(256, 32, 1);
+DECL_FATTN_MMA_F16_CASE(64, 64, 32, 1);
+DECL_FATTN_MMA_F16_CASE(80, 80, 32, 1);
+DECL_FATTN_MMA_F16_CASE(96, 96, 32, 1);
+DECL_FATTN_MMA_F16_CASE(112, 112, 32, 1);
+DECL_FATTN_MMA_F16_CASE(128, 128, 32, 1);
+DECL_FATTN_MMA_F16_CASE(256, 256, 32, 1);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 32, 2);
-DECL_FATTN_MMA_F16_CASE(80, 32, 2);
-DECL_FATTN_MMA_F16_CASE(96, 32, 2);
-DECL_FATTN_MMA_F16_CASE(112, 32, 2);
-DECL_FATTN_MMA_F16_CASE(128, 32, 2);
-DECL_FATTN_MMA_F16_CASE(256, 32, 2);
+DECL_FATTN_MMA_F16_CASE(64, 64, 32, 2);
+DECL_FATTN_MMA_F16_CASE(80, 80, 32, 2);
+DECL_FATTN_MMA_F16_CASE(96, 96, 32, 2);
+DECL_FATTN_MMA_F16_CASE(112, 112, 32, 2);
+DECL_FATTN_MMA_F16_CASE(128, 128, 32, 2);
+DECL_FATTN_MMA_F16_CASE(256, 256, 32, 2);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 4, 2);
-DECL_FATTN_MMA_F16_CASE(80, 4, 2);
-DECL_FATTN_MMA_F16_CASE(96, 4, 2);
-DECL_FATTN_MMA_F16_CASE(112, 4, 2);
-DECL_FATTN_MMA_F16_CASE(128, 4, 2);
-DECL_FATTN_MMA_F16_CASE(256, 4, 2);
+DECL_FATTN_MMA_F16_CASE(64, 64, 4, 2);
+DECL_FATTN_MMA_F16_CASE(80, 80, 4, 2);
+DECL_FATTN_MMA_F16_CASE(96, 96, 4, 2);
+DECL_FATTN_MMA_F16_CASE(112, 112, 4, 2);
+DECL_FATTN_MMA_F16_CASE(128, 128, 4, 2);
+DECL_FATTN_MMA_F16_CASE(256, 256, 4, 2);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 4, 4);
-DECL_FATTN_MMA_F16_CASE(80, 4, 4);
-DECL_FATTN_MMA_F16_CASE(96, 4, 4);
-DECL_FATTN_MMA_F16_CASE(112, 4, 4);
-DECL_FATTN_MMA_F16_CASE(128, 4, 4);
-DECL_FATTN_MMA_F16_CASE(256, 4, 4);
+DECL_FATTN_MMA_F16_CASE(64, 64, 4, 4);
+DECL_FATTN_MMA_F16_CASE(80, 80, 4, 4);
+DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
+DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
+DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
+DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 4, 8);
-DECL_FATTN_MMA_F16_CASE(80, 4, 8);
-DECL_FATTN_MMA_F16_CASE(96, 4, 8);
-DECL_FATTN_MMA_F16_CASE(112, 4, 8);
-DECL_FATTN_MMA_F16_CASE(128, 4, 8);
-DECL_FATTN_MMA_F16_CASE(256, 4, 8);
+DECL_FATTN_MMA_F16_CASE(64, 64, 4, 8);
+DECL_FATTN_MMA_F16_CASE(80, 80, 4, 8);
+DECL_FATTN_MMA_F16_CASE(96, 96, 4, 8);
+DECL_FATTN_MMA_F16_CASE(112, 112, 4, 8);
+DECL_FATTN_MMA_F16_CASE(128, 128, 4, 8);
+DECL_FATTN_MMA_F16_CASE(256, 256, 4, 8);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 64, 1);
-DECL_FATTN_MMA_F16_CASE(80, 64, 1);
-DECL_FATTN_MMA_F16_CASE(96, 64, 1);
-DECL_FATTN_MMA_F16_CASE(112, 64, 1);
-DECL_FATTN_MMA_F16_CASE(128, 64, 1);
-DECL_FATTN_MMA_F16_CASE(256, 64, 1);
+DECL_FATTN_MMA_F16_CASE(64, 64, 64, 1);
+DECL_FATTN_MMA_F16_CASE(80, 80, 64, 1);
+DECL_FATTN_MMA_F16_CASE(96, 96, 64, 1);
+DECL_FATTN_MMA_F16_CASE(112, 112, 64, 1);
+DECL_FATTN_MMA_F16_CASE(128, 128, 64, 1);
+DECL_FATTN_MMA_F16_CASE(256, 256, 64, 1);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 8, 1);
-DECL_FATTN_MMA_F16_CASE(80, 8, 1);
-DECL_FATTN_MMA_F16_CASE(96, 8, 1);
-DECL_FATTN_MMA_F16_CASE(112, 8, 1);
-DECL_FATTN_MMA_F16_CASE(128, 8, 1);
-DECL_FATTN_MMA_F16_CASE(256, 8, 1);
+DECL_FATTN_MMA_F16_CASE(64, 64, 8, 1);
+DECL_FATTN_MMA_F16_CASE(80, 80, 8, 1);
+DECL_FATTN_MMA_F16_CASE(96, 96, 8, 1);
+DECL_FATTN_MMA_F16_CASE(112, 112, 8, 1);
+DECL_FATTN_MMA_F16_CASE(128, 128, 8, 1);
+DECL_FATTN_MMA_F16_CASE(256, 256, 8, 1);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 8, 2);
-DECL_FATTN_MMA_F16_CASE(80, 8, 2);
-DECL_FATTN_MMA_F16_CASE(96, 8, 2);
-DECL_FATTN_MMA_F16_CASE(112, 8, 2);
-DECL_FATTN_MMA_F16_CASE(128, 8, 2);
-DECL_FATTN_MMA_F16_CASE(256, 8, 2);
+DECL_FATTN_MMA_F16_CASE(64, 64, 8, 2);
+DECL_FATTN_MMA_F16_CASE(80, 80, 8, 2);
+DECL_FATTN_MMA_F16_CASE(96, 96, 8, 2);
+DECL_FATTN_MMA_F16_CASE(112, 112, 8, 2);
+DECL_FATTN_MMA_F16_CASE(128, 128, 8, 2);
+DECL_FATTN_MMA_F16_CASE(256, 256, 8, 2);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 8, 4);
-DECL_FATTN_MMA_F16_CASE(80, 8, 4);
-DECL_FATTN_MMA_F16_CASE(96, 8, 4);
-DECL_FATTN_MMA_F16_CASE(112, 8, 4);
-DECL_FATTN_MMA_F16_CASE(128, 8, 4);
-DECL_FATTN_MMA_F16_CASE(256, 8, 4);
+DECL_FATTN_MMA_F16_CASE(64, 64, 8, 4);
+DECL_FATTN_MMA_F16_CASE(80, 80, 8, 4);
+DECL_FATTN_MMA_F16_CASE(96, 96, 8, 4);
+DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4);
+DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4);
+DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu

@@ -2,9 +2,9 @@
 
 #include "../fattn-mma-f16.cuh"
 
-DECL_FATTN_MMA_F16_CASE(64, 8, 8);
-DECL_FATTN_MMA_F16_CASE(80, 8, 8);
-DECL_FATTN_MMA_F16_CASE(96, 8, 8);
-DECL_FATTN_MMA_F16_CASE(112, 8, 8);
-DECL_FATTN_MMA_F16_CASE(128, 8, 8);
-DECL_FATTN_MMA_F16_CASE(256, 8, 8);
+DECL_FATTN_MMA_F16_CASE(64, 64, 8, 8);
+DECL_FATTN_MMA_F16_CASE(80, 80, 8, 8);
+DECL_FATTN_MMA_F16_CASE(96, 96, 8, 8);
+DECL_FATTN_MMA_F16_CASE(112, 112, 8, 8);
+DECL_FATTN_MMA_F16_CASE(128, 128, 8, 8);
+DECL_FATTN_MMA_F16_CASE(256, 256, 8, 8);

ggml/src/ggml-cuda/template-instances/generate_cu_files.py

@@ -18,7 +18,7 @@ SOURCE_FATTN_MMA_START = """// This file has been autogenerated by generate_cu_f
 
 """
 
-SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size}, {ncols1}, {ncols2});\n"
+SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size_kq}, {head_size_v}, {ncols1}, {ncols2});\n"
 
 TYPES_MMQ = [
     "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0",
@@ -57,18 +57,21 @@ for vkq_size in [16, 32]:
                 with open(f"fattn-vec-f{vkq_size}-instance-hs{head_size}-{get_short_name(type_k)}-{get_short_name(type_v)}.cu", "w") as f:
                     f.write(SOURCE_FATTN_VEC.format(vkq_size=vkq_size, head_size=head_size, type_k=type_k, type_v=type_v))
 
-for ncols in [8, 16, 32, 64, 128]:
-    for ncols2 in [1, 2, 4, 8]:
+for ncols in [8, 16, 32, 64]:
+    for ncols2 in [1, 2, 4, 8, 16]:
+        if ncols2 > ncols:
+            continue
         ncols1 = ncols // ncols2
-        if ncols == 128:
-            continue  # Too much register pressure.
         with open(f"fattn-mma-f16-instance-ncols1_{ncols1}-ncols2_{ncols2}.cu", "w") as f:
             f.write(SOURCE_FATTN_MMA_START)
 
-            for head_size in [64, 80, 96, 112, 128, 256]:
-                if ncols == 128 and head_size == 256:
-                    continue  # Needs too much shared memory.
-                f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size=head_size))
+            for head_size_kq in [64, 80, 96, 112, 128, 256, 576]:
+                if head_size_kq != 576 and ncols2 == 16:
+                    continue
+                if head_size_kq == 576 and ncols2 != 16:
+                    continue
+                head_size_v = head_size_kq if head_size_kq != 576 else 512
+                f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size_kq=head_size_kq, head_size_v=head_size_v))
 
 for type in TYPES_MMQ:
     with open(f"mmq-instance-{get_short_name(type)}.cu", "w") as f:

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -299,21 +299,42 @@ typedef struct {
 } ggml_metal_kargs_mul_mv_ext;
 
 typedef struct {
-    int32_t  nei0;
-    int32_t  nei1;
-    uint64_t nbi1;
+    int32_t  ne10;
+    int32_t  ne11;  // n_expert_used (bcast)
+    uint64_t nb11;
+    uint64_t nb12;
+    int32_t  neh11; // n_tokens
+    uint64_t nbh11;
+    int32_t  ne20;  // n_expert_used
+    uint64_t nb21;
+} ggml_metal_kargs_mul_mm_id_map0;
+
+typedef struct {
+    int32_t  ne20; // n_expert_used
+    int32_t  neh0;
+    int32_t  neh1;
+    uint64_t nbh1;
+    uint64_t nbh2;
+    int32_t  ne0;
+    uint64_t nb1;
+    uint64_t nb2;
+} ggml_metal_kargs_mul_mm_id_map1;
+
+typedef struct {
     int32_t  ne00;
     int32_t  ne02;
     uint64_t nb01;
     uint64_t nb02;
-    int32_t  ne11;
-    int32_t  ne12;
-    int32_t  ne13;
-    uint64_t nb10;
-    uint64_t nb11;
-    uint64_t nb12;
-    int32_t  ne0;
-    int32_t  ne1;
+    uint64_t nb03;
+    int32_t  neh12;
+    uint64_t nbh10;
+    uint64_t nbh11;
+    uint64_t nbh12;
+    uint64_t nbh13;
+    int32_t  neh0;
+    int32_t  neh1;
+    int16_t  r2;
+    int16_t  r3;
 } ggml_metal_kargs_mul_mm_id;
 
 typedef struct {

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

@@ -306,28 +306,30 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP1_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F16,
     GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16,
     GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32,
@@ -650,7 +652,8 @@ static void ggml_metal_mem_pool_reset(struct ggml_metal_mem_pool * mem_pool) {
     }
 
     if (mem_pool->heaps_to_remove.count > 0) {
-        for (NSUInteger i = 0; i < [mem_pool->heaps_to_remove count]; i++) {
+        // remove in reverse order
+        for (NSUInteger i = [mem_pool->heaps_to_remove count] - 1; ; --i) {
             NSUInteger index = [[mem_pool->heaps_to_remove objectAtIndex:i] intValue];
             ggml_metal_heap_ptr * ptr = [mem_pool->heaps objectAtIndex:index];
 
@@ -659,6 +662,10 @@ static void ggml_metal_mem_pool_reset(struct ggml_metal_mem_pool * mem_pool) {
 
             [mem_pool->heaps removeObjectAtIndex:index];
             [ptr release];
+
+            if (i == 0) {
+                break;
+            }
         }
 
         [mem_pool->heaps_to_remove removeAllObjects];
@@ -672,7 +679,7 @@ static void ggml_metal_mem_pool_clear(struct ggml_metal_mem_pool * mem_pool) {
 }
 
 static id<MTLBuffer> ggml_metal_mem_pool_alloc(struct ggml_metal_mem_pool * mem_pool, size_t size) {
-    const size_t alignment = 32;
+    const size_t alignment = 256;
 
     const size_t size_aligned = GGML_PAD(size, alignment);
 
@@ -1242,28 +1249,30 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,                mul_mm_iq1_m_f32,                has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,               mul_mm_iq4_nl_f32,               has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,               mul_mm_iq4_xs_f32,               has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,               mul_mm_id_f32_f32,               has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,               mul_mm_id_f16_f32,               has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32,              mul_mm_id_bf16_f32,              has_simdgroup_mm && use_bfloat);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,              mul_mm_id_q4_0_f32,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,              mul_mm_id_q4_1_f32,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,              mul_mm_id_q5_0_f32,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,              mul_mm_id_q5_1_f32,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,              mul_mm_id_q8_0_f32,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,              mul_mm_id_q2_K_f32,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,              mul_mm_id_q3_K_f32,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,              mul_mm_id_q4_K_f32,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,              mul_mm_id_q5_K_f32,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,              mul_mm_id_q6_K_f32,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,           mul_mm_id_iq2_xxs_f32,           has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,            mul_mm_id_iq2_xs_f32,            has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,           mul_mm_id_iq3_xxs_f32,           has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,             mul_mm_id_iq3_s_f32,             has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,             mul_mm_id_iq2_s_f32,             has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,             mul_mm_id_iq1_s_f32,             has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,             mul_mm_id_iq1_m_f32,             has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,            mul_mm_id_iq4_nl_f32,            has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,            mul_mm_id_iq4_xs_f32,            has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16,              mul_mm_id_map0_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP1_F32,              mul_mm_id_map1_f32,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F16,               mul_mm_id_f32_f16,               has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F16,               mul_mm_id_f16_f16,               has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F16,              mul_mm_id_bf16_f16,              has_simdgroup_mm && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F16,              mul_mm_id_q4_0_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F16,              mul_mm_id_q4_1_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16,              mul_mm_id_q5_0_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16,              mul_mm_id_q5_1_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16,              mul_mm_id_q8_0_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16,              mul_mm_id_q2_K_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16,              mul_mm_id_q3_K_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16,              mul_mm_id_q4_K_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F16,              mul_mm_id_q5_K_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F16,              mul_mm_id_q6_K_f16,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F16,           mul_mm_id_iq2_xxs_f16,           has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F16,            mul_mm_id_iq2_xs_f16,            has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F16,           mul_mm_id_iq3_xxs_f16,           has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F16,             mul_mm_id_iq3_s_f16,             has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F16,             mul_mm_id_iq2_s_f16,             has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F16,             mul_mm_id_iq1_s_f16,             has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F16,             mul_mm_id_iq1_m_f16,             has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F16,            mul_mm_id_iq4_nl_f16,            has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F16,            mul_mm_id_iq4_xs_f16,            has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32,                   rope_norm_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16,                   rope_norm_f16,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32,                   rope_neox_f32,                   true);
@@ -2999,7 +3008,7 @@ static bool ggml_metal_encode_node(
                     [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
 
                     [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                    [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                    [encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                 } else {
                     id<MTLComputePipelineState> pipeline = nil;
 
@@ -3219,8 +3228,6 @@ static bool ggml_metal_encode_node(
             } break;
         case GGML_OP_MUL_MAT_ID:
             {
-                const int n_as = src0->ne[2];
-
                 // src2 = ids
                 const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t);
 
@@ -3234,24 +3241,21 @@ static bool ggml_metal_encode_node(
                 GGML_ASSERT(ne03 == 1);
                 GGML_ASSERT(ne13 == 1);
 
+                const uint32_t r2 = 1;
+                const uint32_t r3 = 1;
+
                 // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                 // to the matrix-vector kernel
                 // ne20 = n_used_experts
-                // ne21 = n_rows
-                const int dst_rows = ne20*ne21;
-                const int dst_rows_min = n_as;
-                const int dst_rows_max = (device.maxThreadgroupMemoryLength/2 - 8192)/4;
-
-                // max size of the rowids array in the kernel shared buffer
-                //GGML_ASSERT(dst_rows <= dst_rows_max);
+                // ne21 = n_rows (batch size)
+                const int ne21_mm_id_min = 32;
 
                 // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                 // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
                 if ([device supportsFamily:MTLGPUFamilyApple7] &&
                         ne00 % 32 == 0 && ne00 >= 64 &&
-                        //ne01 / ne02 >= 512 &&    // NOTE: this is based on Mixtral shapes, might need adjustments
-                        dst_rows >  dst_rows_min &&
-                        dst_rows <= dst_rows_max) {
+                        (ne21 >= ne21_mm_id_min)) {
+                    GGML_ASSERT(ne00 % 4 == 0);
 
                     // some Metal matrix data types require aligned pointers
                     // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
@@ -3262,62 +3266,169 @@ static bool ggml_metal_encode_node(
                         default: break;
                     }
 
-                    id<MTLComputePipelineState> pipeline = nil;
+                    const int64_t neh10 = ne10; // n_embd
+                    const int64_t neh11 = ne21; // n_tokens
+                    const int64_t neh12 = ne02; // n_expert
 
-                    switch (src0->type) {
-                        case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32    ].pipeline; break;
-                        case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32    ].pipeline; break;
-                        case GGML_TYPE_BF16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32   ].pipeline; break;
-                        case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32   ].pipeline; break;
-                        case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32   ].pipeline; break;
-                        case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32   ].pipeline; break;
-                        case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32   ].pipeline; break;
-                        case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32   ].pipeline; break;
-                        case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32   ].pipeline; break;
-                        case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32   ].pipeline; break;
-                        case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32   ].pipeline; break;
-                        case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32   ].pipeline; break;
-                        case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32   ].pipeline; break;
-                        case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
-                        case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
-                        case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
-                        case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32  ].pipeline; break;
-                        case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32  ].pipeline; break;
-                        case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32  ].pipeline; break;
-                        case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32  ].pipeline; break;
-                        case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
-                        case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
-                        default: GGML_ABORT("MUL_MAT_ID not implemented");
+                    const uint64_t nbh10 = ggml_type_size(GGML_TYPE_F16);
+                    const uint64_t nbh11 = nbh10*neh10;
+                    const uint64_t nbh12 = nbh11*neh11;
+                    const uint64_t nbh13 = nbh12*neh12;
+
+                    const size_t s_src1 = ggml_type_size(GGML_TYPE_F16)*neh10*neh11*neh12;
+                    id<MTLBuffer> h_src1 = ggml_metal_mem_pool_alloc(mem_pool, s_src1);
+                    if (!h_src1) {
+                        GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_src1);
+                        return false;
                     }
 
-                    ggml_metal_kargs_mul_mm_id args = {
-                        /*.nei0 =*/ ne20,
-                        /*.nei1 =*/ ne21,
-                        /*.nbi1 =*/ nb21,
-                        /*.ne00 =*/ ne00,
-                        /*.ne02 =*/ ne02,
-                        /*.nb01 =*/ nb01,
-                        /*.nb02 =*/ nb02,
-                        /*.ne11 =*/ ne11,
-                        /*.ne12 =*/ ne12,
-                        /*.ne13 =*/ ne13,
-                        /*.nb10 =*/ nb10,
-                        /*.nb11 =*/ nb11,
-                        /*.nb12 =*/ nb12,
-                        /*.ne0  =*/ ne0,
-                        /*.ne1  =*/ ne1,
-                    };
+                    const int64_t neh0 = ne0;
+                    const int64_t neh1 = ne21;
+                    const int64_t neh2 = ne02;
 
-                    [encoder setComputePipelineState:pipeline];
-                    [encoder setBytes:&args    length:sizeof(args) atIndex:0];
-                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
-                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
-                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
-                    [encoder setBuffer:id_src2 offset:offs_src2    atIndex:4];
+                    const uint64_t nbh0 = ggml_type_size(GGML_TYPE_F32);
+                    const uint64_t nbh1 = nbh0*neh0;
+                    const uint64_t nbh2 = nbh1*neh1;
+                  //const uint64_t nbh3 = nbh2*neh2;
 
-                    [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0];
+                    const size_t s_dst = ggml_type_size(GGML_TYPE_F32)*neh0*neh1*neh2;
+                    id<MTLBuffer> h_dst = ggml_metal_mem_pool_alloc(mem_pool, s_dst);
+                    if (!h_dst) {
+                        GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_dst);
+                        return false;
+                    }
 
-                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                    // tokens per expert
+                    const size_t s_tpe = ggml_type_size(GGML_TYPE_I32)*ne02;
+                    id<MTLBuffer> h_tpe = ggml_metal_mem_pool_alloc(mem_pool, s_tpe);
+                    if (!h_tpe) {
+                        GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_tpe);
+                        return false;
+                    }
+
+                    // id map
+                    // [n_expert_used, n_tokens]
+                    const size_t s_ids = ggml_type_size(GGML_TYPE_I32)*ne20*ne21;
+                    id<MTLBuffer> h_ids = ggml_metal_mem_pool_alloc(mem_pool, s_ids);
+                    if (!h_ids) {
+                        GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_ids);
+                        return false;
+                    }
+
+                    {
+                        const int nth = MIN(1024, ne10/4);
+
+                        ggml_metal_kargs_mul_mm_id_map0 args = {
+                            ne10,
+                            ne11,  // n_expert_used (bcast)
+                            nb11,
+                            nb12,
+                            neh11, // n_tokens
+                            nbh11,
+                            ne20,  // n_expert_used
+                            nb21,
+                        };
+
+                        id<MTLComputePipelineState> pipeline = nil;
+
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16].pipeline;
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                        [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
+                        [encoder setBuffer:id_src2 offset:offs_src2    atIndex:2];
+                        [encoder setBuffer: h_src1 offset:0            atIndex:3];
+                        [encoder setBuffer: h_tpe  offset:0            atIndex:4];
+                        [encoder setBuffer: h_ids  offset:0            atIndex:5];
+
+                        [encoder dispatchThreadgroups:MTLSizeMake(ne02, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                    }
+
+                    {
+                        id<MTLComputePipelineState> pipeline = nil;
+
+                        switch (src0->type) {
+                            case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F16    ].pipeline; break;
+                            case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F16    ].pipeline; break;
+                            case GGML_TYPE_BF16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F16   ].pipeline; break;
+                            case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F16   ].pipeline; break;
+                            case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F16   ].pipeline; break;
+                            case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16   ].pipeline; break;
+                            case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16   ].pipeline; break;
+                            case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16   ].pipeline; break;
+                            case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16   ].pipeline; break;
+                            case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16   ].pipeline; break;
+                            case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16   ].pipeline; break;
+                            case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F16   ].pipeline; break;
+                            case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F16   ].pipeline; break;
+                            case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F16].pipeline; break;
+                            case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F16 ].pipeline; break;
+                            case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F16].pipeline; break;
+                            case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F16  ].pipeline; break;
+                            case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F16  ].pipeline; break;
+                            case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F16  ].pipeline; break;
+                            case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F16  ].pipeline; break;
+                            case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F16 ].pipeline; break;
+                            case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F16 ].pipeline; break;
+                            default: GGML_ABORT("MUL_MAT_ID not implemented");
+                        }
+
+                        ggml_metal_kargs_mul_mm_id args = {
+                            /*.ne00  =*/ ne00,
+                            /*.ne02  =*/ ne02,
+                            /*.nb01  =*/ nb01,
+                            /*.nb02  =*/ nb02,
+                            /*.nb03  =*/ nb03,
+                            /*.neh12 =*/ neh12,
+                            /*.nbh10 =*/ nbh10,
+                            /*.nbh11 =*/ nbh11,
+                            /*.nbh12 =*/ nbh12,
+                            /*.nbh13 =*/ nbh13,
+                            /*.neh0  =*/ neh0,
+                            /*.neh1  =*/ neh1,
+                            /*.r2    =*/ r2,
+                            /*.r3    =*/ r3,
+                        };
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                        [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                        [encoder setBuffer: h_src1 offset:0            atIndex:2];
+                        [encoder setBuffer: h_tpe  offset:0            atIndex:3];
+                        [encoder setBuffer: h_dst  offset:0            atIndex:4];
+
+                        [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, ne02) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                    }
+
+                    {
+                        GGML_ASSERT(ne0 % 4 == 0);
+
+                        const int nth = MIN(1024, ne0/4);
+
+                        ggml_metal_kargs_mul_mm_id_map1 args = {
+                            ne20, // n_expert_used
+                            neh0,
+                            neh1,
+                            nbh1,
+                            nbh2,
+                            ne0,
+                            nb1,
+                            nb2,
+                        };
+
+                        id<MTLComputePipelineState> pipeline = nil;
+
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP1_F32].pipeline;
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBytes:&args   length:sizeof(args) atIndex:0];
+                        [encoder setBuffer: h_dst offset:0            atIndex:1];
+                        [encoder setBuffer: h_ids offset:0            atIndex:2];
+                        [encoder setBuffer:id_dst offset:offs_dst     atIndex:3];
+
+                        [encoder dispatchThreadgroups:MTLSizeMake(ne20, ne21, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                    }
                 } else {
                     id<MTLComputePipelineState> pipeline = nil;
 
@@ -3511,7 +3622,7 @@ static bool ggml_metal_encode_node(
                     [encoder setBuffer:id_src2 offset:offs_src2 atIndex:4];
 
                     const int64_t _ne1 = 1;
-                    const int64_t ne123 = dst_rows;
+                    const int64_t ne123 = ne20*ne21;
 
                     if (smem > 0) {
                         [encoder setThreadgroupMemoryLength:smem atIndex:0];

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

@@ -6336,127 +6336,219 @@ kernel void kernel_mul_mm(
     }
 }
 
-// same as kernel_mul_mm_impl, but src1 and dst are accessed via indices stored in rowids
-// TODO: this kernel needs to be reimplemented from scratch for better performance
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-void kernel_mul_mm_id_impl(
-        int32_t  ne00,
-        int32_t  ne02,
-        uint64_t nb01,
-        uint64_t nb02,
-        int32_t  ne11,
-        int32_t  ne12,
-        uint64_t nb10,
-        uint64_t nb11,
-        uint64_t nb12,
-        int32_t  ne0,
-        int32_t  ne1,
-        int64_t  ne0ne1,
-        device   const char * src0,
-        device   const char * src1,
-        threadgroup ushort2 * rowids,
-        device         char * dst,
-        threadgroup    char * shmem,
+template<typename T4>
+kernel void kernel_mul_mm_id_map0(
+        constant ggml_metal_kargs_mul_mm_id_map0 & args,
+        device  const char * src1,
+        device  const char * src2,
+        device        char * hsrc1,
+        device        char * htpe,
+        device        char * hids,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ide = tgpig[0]; // expert id
+
+    int n_all = 0;
+
+    device int32_t * ids_i32 = (device int32_t *) (hids);
+
+    for (int i21 = 0; i21 < args.neh11; i21++) { // n_tokens
+        device const int32_t * src2_i32 = (device const int32_t *) (src2 + i21*args.nb21);
+
+        for (int i20 = 0; i20 < args.ne20; i20++) { // n_expert_used
+            if (src2_i32[i20] != ide) {
+                continue;
+            }
+
+            device const float4 *  src1_f32x4 = (device const float4 *) ( src1 + i21*args.nb12 + (i20%args.ne11)*args.nb11);
+            device       T4     * hsrc1_f32x4 = (device       T4     *) (hsrc1 + (ide*args.neh11 + n_all)*args.nbh11);
+
+            for (int64_t i00 = tpitg.x; i00 < args.ne10/4; i00 += ntg.x) {
+                hsrc1_f32x4[i00] = (T4) (src1_f32x4[i00]);
+            }
+
+            if (tpitg.x == 0) {
+                ids_i32[i21*args.ne20 + i20] = ide*args.neh11 + n_all;
+            }
+
+            ++n_all;
+        }
+    }
+
+    if (tpitg.x == 0) {
+        device int32_t * tpe_i32 = (device int32_t *) (htpe);
+        tpe_i32[ide] = n_all;
+    }
+}
+
+typedef decltype(kernel_mul_mm_id_map0<half4>) kernel_mul_mm_id_map0_t;
+
+template [[host_name("kernel_mul_mm_id_map0_f16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<half4>;
+
+template<typename T>
+kernel void kernel_mul_mm_id_map1(
+        constant ggml_metal_kargs_mul_mm_id_map1 & args,
+        device  const char * hdst,
+        device  const char * hids,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i20 = tgpig[0]; // used expert
+    const int i21 = tgpig[1]; // token
+
+    device const int32_t * ids_i32    = (device const int32_t *) (hids);
+    device       float4  * dst_f32x4  = (device       float4  *) (dst + i20*args.nb1 + i21*args.nb2);
+
+    const int id = ids_i32[i21*args.ne20 + i20];
+
+    const int ide = id / args.neh1;
+    const int idt = id % args.neh1;
+
+    device const float4 * hdst_f32x4 = (device const float4 *) (hdst + idt*args.nbh1 + ide*args.nbh2);
+
+    for (int64_t i0 = tpitg.x; i0 < args.neh0/4; i0 += ntg.x) {
+        dst_f32x4[i0] = hdst_f32x4[i0];
+    }
+}
+
+typedef decltype(kernel_mul_mm_id_map1<float>) kernel_mul_mm_id_map1_t;
+
+template [[host_name("kernel_mul_mm_id_map1_f32")]] kernel kernel_mul_mm_id_map1_t kernel_mul_mm_id_map1<float>;
+
+template<typename T, typename T4x4, typename simdgroup_T8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
+kernel void kernel_mul_mm_id(
+        constant ggml_metal_kargs_mul_mm_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * tpe,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
         uint3  tgpig[[threadgroup_position_in_grid]],
         ushort tiitg[[thread_index_in_threadgroup]],
         ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    threadgroup half  * sa = (threadgroup half  *)(shmem);
-    threadgroup float * sb = (threadgroup float *)(shmem + 4096);
+    threadgroup T    * sa = (threadgroup T    *)(shmem);
+    threadgroup half * sb = (threadgroup half *)(shmem + 4096);
 
     const int r0 = tgpig.y;
     const int r1 = tgpig.x;
+    const int im = tgpig.z;
 
-    if (r1*BLOCK_SIZE_N >= ne1) return;
+    device const int32_t * tpe_i32 = (device const int32_t *) (tpe);
+
+    const int neh1 = tpe_i32[im];
+
+    if (r1*BLOCK_SIZE_N >= neh1) {
+        return;
+    }
 
     // if this block is of 64x32 shape or smaller
-    short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short n_rows = (args.neh0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.neh0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
+    const short n_cols = (     neh1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (     neh1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
 
     // a thread shouldn't load data outside of the matrix
-    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
+    const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
 
-    simdgroup_half8x8  ma[4];
-    simdgroup_float8x8 mb[2];
+    simdgroup_T8x8     ma[4];
+    simdgroup_half8x8  mb[2];
     simdgroup_float8x8 mc[8];
-    for (int i = 0; i < 8; i++){
+
+    for (short i = 0; i < 8; i++){
         mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
     }
+
     short il = (tiitg % THREAD_PER_ROW);
 
-    ushort offset1 = il/nl;
+    const int i12 = im%args.neh12;
+    const int i13 = im/args.neh12;
 
-    threadgroup const auto & id = rowids[r1 * BLOCK_SIZE_N + thread_col];
+    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const short    offset1 = il/nl;
 
-    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01) + offset1;
-    device const float   * y = (device const float   *)(src1
-        + nb12 * id[1]
-        + nb11 * (id[0] % ne11)
-        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+    device const block_q * x = (device const block_q *)(src0
+        + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
 
-    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
+    device const half   * y = (device const half   *)(src1
+        + args.nbh13*i13
+        + args.nbh12*i12
+        + args.nbh11*(r1*BLOCK_SIZE_N + thread_col)
+        + args.nbh10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
         // load data and store to threadgroup memory
-        half4x4 temp_a;
+        T4x4 temp_a;
         dequantize_func(x, il, temp_a);
+
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        for (int i = 0; i < 16; i++) {
-            *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
-            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
-            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
+        #pragma unroll(16)
+        for (short i = 0; i < 16; i++) {
+            *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
+            +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
+            +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
         }
 
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
+        *(threadgroup half2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = *((device half2x4 *) y);
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
-        x  = (il < 2) ? x + (2+nl-1)/nl : x;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
         y += BLOCK_SIZE_K;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
-        threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
+        threadgroup const T    * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
+        threadgroup const half * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
 
-        #pragma unroll(BLOCK_SIZE_K/8)
-        for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
+        #pragma unroll(4)
+        for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
             #pragma unroll(4)
-            for (int i = 0; i < 4; i++) {
+            for (short i = 0; i < 4; i++) {
                 simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
             }
+
             simdgroup_barrier(mem_flags::mem_none);
+
             #pragma unroll(2)
-            for (int i = 0; i < 2; i++) {
+            for (short i = 0; i < 2; i++) {
                 simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
             }
 
-            lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
-            lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
-
             #pragma unroll(8)
-            for (int i = 0; i < 8; i++){
+            for (short i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
+
+            lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE;
+            lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE;
         }
     }
 
-    {
+    if ((r0 + 1) * BLOCK_SIZE_M <= args.neh0 && (r1 + 1) * BLOCK_SIZE_N <= neh1) {
+        device float * C = (device float *) dst +
+            (BLOCK_SIZE_M * r0 + 32*(sgitg &  1)) + \
+            (BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.neh0 + im*args.neh1*args.neh0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.neh0 * (i/4), args.neh0);
+        }
+    } else {
+        // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
         threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                      + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
-        for (int i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
+                                     + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
         }
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         if (sgitg == 0) {
             for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                threadgroup const auto & jid = rowids[r1 * BLOCK_SIZE_N + j];
-                int64_t joff = jid[0]*ne0 + jid[1]*ne0ne1;
-
-                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + joff;
+                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*args.neh0 + im*args.neh1*args.neh0;
                 device float4 * D4 = (device float4 *) D;
 
                 threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
@@ -6476,66 +6568,6 @@ void kernel_mul_mm_id_impl(
     }
 }
 
-template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-kernel void kernel_mul_mm_id(
-        constant ggml_metal_kargs_mul_mm_id & args,
-        device const char * src0s,
-        device const char * src1,
-        device       char * dst,
-        device const char * ids,
-        threadgroup  char * shmem [[threadgroup(0)]],
-        uint3  tgpig[[threadgroup_position_in_grid]],
-        ushort tiitg[[thread_index_in_threadgroup]],
-        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const int32_t i02 = tgpig.z;
-
-    tgpig.z = 0;
-
-    device const char * src0 = src0s + i02*args.nb02;
-
-    // row indices
-    threadgroup ushort2 * rowids = (threadgroup ushort2 *)(shmem + 8192);
-
-    // TODO: parallelize this loop
-    int32_t _ne1 = 0;
-    for (ushort ii1 = 0; ii1 < args.nei1; ii1++) {
-        for (ushort ii0 = 0; ii0 < args.nei0; ii0++) {
-            int32_t id = ((device int32_t *) (ids + ii1*args.nbi1))[ii0];
-            if (id == i02) {
-                if (tiitg == 0) {
-                    rowids[_ne1] = ushort2(ii0, ii1);
-                }
-                _ne1++;
-            }
-        }
-    }
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    kernel_mul_mm_id_impl<block_q, nl, dequantize_func>(
-        args.ne00,
-        args.ne02,
-        args.nb01,
-        args.nb02,
-        args.ne11,
-        args.ne12,
-        args.nb10,
-        args.nb11,
-        args.nb12,
-        args.ne0,
-        _ne1,
-        (int64_t)args.ne0*args.ne1,
-        src0,
-        src1,
-        rowids,
-        dst,
-        shmem,
-        tgpig,
-        tiitg,
-        sgitg);
-}
-
 #define QK_NL 16
 
 //
@@ -6576,63 +6608,64 @@ template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get
 // matrix-matrix multiplication
 //
 
-typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, float4x4, 1, dequantize_f32>) mat_mm_t;
+typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, float4x4, 1, dequantize_f32>) mul_mm_t;
 
-template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32>;
-template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16>;
 #if defined(GGML_METAL_USE_BF16)
-template [[host_name("kernel_mul_mm_bf16_f32")]]    kernel mat_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16>;
+template [[host_name("kernel_mul_mm_bf16_f32")]]    kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16>;
 #endif
-template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0>;
-template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1>;
-template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0>;
-template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K>;
-template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
-template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
-template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
-template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s>;
-template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s>;
-template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s>;
-template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m>;
-template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl>;
-template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
 
 //
 // indirect matrix-matrix multiplication
 //
 
-typedef decltype(kernel_mul_mm_id<float4x4, 1, dequantize_f32>) mat_mm_id_t;
+typedef decltype(kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, float4x4, 1, dequantize_f32>) mul_mm_id;
 
-template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mat_mm_id_t kernel_mul_mm_id<float4x4,      1,     dequantize_f32>;
-template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mat_mm_id_t kernel_mul_mm_id<half4x4,       1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_id_f32_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_id_f16_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16>;
 #if defined(GGML_METAL_USE_BF16)
-template [[host_name("kernel_mul_mm_id_bf16_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<bfloat4x4,     1,     dequantize_bf16>;
+template [[host_name("kernel_mul_mm_id_bf16_f16")]]    kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16>;
 #endif
-template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0,    2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1,    2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0,    2,     dequantize_q5_0>;
-template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1,    2,     dequantize_q5_1>;
-template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0,    2,     dequantize_q8_0>;
-template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K,    QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K,    QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K,    QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K,    QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K,    QK_NL, dequantize_q6_K>;
-template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
-template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
-template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
-template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_s,   QK_NL, dequantize_iq3_s>;
-template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_s,   QK_NL, dequantize_iq2_s>;
-template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s,   QK_NL, dequantize_iq1_s>;
-template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_m,   QK_NL, dequantize_iq1_m>;
-template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl,  2,     dequantize_iq4_nl>;
-template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+template [[host_name("kernel_mul_mm_id_q4_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_id_q4_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_id_q5_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_id_q5_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_id_q8_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_id_q2_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_id_q3_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_id_q4_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_id_q5_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_id_q6_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+
 
 //
 // matrix-vector multiplication

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -151,6 +151,12 @@ struct rpc_msg_buffer_clear_req {
     uint8_t value;
 };
 
+struct rpc_msg_set_tensor_hash_req {
+    rpc_tensor tensor;
+    uint64_t offset;
+    uint64_t hash;
+};
+
 struct rpc_msg_set_tensor_hash_rsp {
     uint8_t result;
 };
@@ -548,15 +554,12 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     rpc_tensor rpc_tensor = serialize_tensor(tensor);
     if (size > HASH_THRESHOLD) {
-        // input serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes)
-        size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + sizeof(uint64_t);
-        std::vector<uint8_t> input(input_size, 0);
-        uint64_t hash = fnv_hash((const uint8_t*)data, size);
-        memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
-        memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
-        memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), &hash, sizeof(hash));
+        rpc_msg_set_tensor_hash_req request;
+        request.tensor = rpc_tensor;
+        request.offset = offset;
+        request.hash = fnv_hash((const uint8_t*)data, size);
         rpc_msg_set_tensor_hash_rsp response;
-        bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, input.data(), input.size(), &response, sizeof(response));
+        bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, &request, sizeof(request), &response, sizeof(response));
         GGML_ASSERT(status);
         if (response.result) {
             // the server has the same data, no need to send it
@@ -864,7 +867,7 @@ public:
     bool free_buffer(const rpc_msg_free_buffer_req & request);
     bool buffer_clear(const rpc_msg_buffer_clear_req & request);
     bool set_tensor(const std::vector<uint8_t> & input);
-    bool set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set_tensor_hash_rsp & response);
+    bool set_tensor_hash(const rpc_msg_set_tensor_hash_req & request, rpc_msg_set_tensor_hash_rsp & response);
     bool get_tensor(const rpc_msg_get_tensor_req & request, std::vector<uint8_t> & response);
     bool copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response);
     bool graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph_compute_rsp & response);
@@ -1101,18 +1104,10 @@ bool rpc_server::get_cached_file(uint64_t hash, std::vector<uint8_t> & data) {
     return true;
 }
 
-bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set_tensor_hash_rsp & response)
+bool rpc_server::set_tensor_hash(const rpc_msg_set_tensor_hash_req & request, rpc_msg_set_tensor_hash_rsp & response)
 {
-    // serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes) |
-    if (input.size() != sizeof(rpc_tensor) + 16) {
-        return false;
-    }
-    const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
-    uint64_t offset;
-    memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
-    const uint64_t * hash = (const uint64_t *)(input.data() + sizeof(rpc_tensor) + sizeof(offset));
     std::vector<uint8_t> cached_file;
-    if (!get_cached_file(*hash, cached_file)) {
+    if (!get_cached_file(request.hash, cached_file)) {
         response.result = 0;
         return true;
     }
@@ -1125,25 +1120,28 @@ bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set
     ggml_context_ptr ctx_ptr { ggml_init(params) };
     GGML_ASSERT(ctx_ptr != nullptr);
     ggml_context * ctx = ctx_ptr.get();
-    ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
+    ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
     if (tensor == nullptr) {
         GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__);
         return false;
     }
-    GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size, *hash);
+    GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n",
+        __func__, (void*)tensor->buffer, tensor->data, request.offset, size, request.hash);
 
     // sanitize tensor->data
     {
         const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
         const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
 
-        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+        if (request.tensor.data + request.offset < p0
+         || request.tensor.data + request.offset >= p1
+         || size > (p1 - request.tensor.data - request.offset)) {
             GGML_LOG_ERROR("[%s] tensor data region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%zu, hash=0x%" PRIx64 ") out of buffer bounds [0x%zx, 0x%zx)\n",
-                           __func__, in_tensor->data, offset, size, *hash, p0, p1);
+                           __func__, request.tensor.data, request.offset, size, request.hash, p0, p1);
             return false;
         }
     }
-    ggml_backend_tensor_set(tensor, cached_file.data(), offset, size);
+    ggml_backend_tensor_set(tensor, cached_file.data(), request.offset, size);
     response.result = 1;
     return true;
 }
@@ -1503,12 +1501,12 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
                 break;
             }
             case RPC_CMD_SET_TENSOR_HASH: {
-                std::vector<uint8_t> input;
-                if (!recv_msg(sockfd, input)) {
+                rpc_msg_set_tensor_hash_req request;
+                if (!recv_msg(sockfd, &request, sizeof(request))) {
                     return;
                 }
                 rpc_msg_set_tensor_hash_rsp response;
-                if (!server.set_tensor_hash(input, response)) {
+                if (!server.set_tensor_hash(request, response)) {
                     return;
                 }
                 if (!send_msg(sockfd, &response, sizeof(response))) {

ggml/src/ggml-sycl/backend.hpp

@@ -14,23 +14,24 @@
 #define GGML_SYCL_BACKEND_HPP
 
 #include "binbcast.hpp"
-#include "concat.hpp"
 #include "common.hpp"
+#include "concat.hpp"
 #include "conv.hpp"
 #include "convert.hpp"
+#include "cpy.hpp"
 #include "dequantize.hpp"
 #include "dmmv.hpp"
+#include "element_wise.hpp"
+#include "gla.hpp"
+#include "im2col.hpp"
 #include "mmq.hpp"
 #include "mmvq.hpp"
-#include "rope.hpp"
 #include "norm.hpp"
+#include "outprod.hpp"
+#include "quants.hpp"
+#include "rope.hpp"
 #include "softmax.hpp"
 #include "tsembd.hpp"
-#include "im2col.hpp"
 #include "wkv.hpp"
-#include "outprod.hpp"
-#include "element_wise.hpp"
-#include "cpy.hpp"
-#include "gla.hpp"
 
-#endif // GGML_SYCL_BACKEND_HPP
+#endif  // GGML_SYCL_BACKEND_HPP

ggml/src/ggml-sycl/common.hpp

@@ -42,6 +42,7 @@ void ggml_sycl_host_free(void* ptr);
 
 extern int g_ggml_sycl_debug;
 extern int g_ggml_sycl_disable_optimize;
+extern int g_ggml_sycl_prioritize_dmmv;
 
 #define GGML_SYCL_DEBUG(...)        \
   do {                              \

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -49,6 +49,7 @@ static bool g_sycl_loaded = false;
 int g_ggml_sycl_debug = 0;
 int g_ggml_sycl_disable_optimize = 0;
 int g_ggml_sycl_disable_graph = 0;
+int g_ggml_sycl_prioritize_dmmv = 0;
 
 static ggml_sycl_device_info ggml_sycl_init() {
     ggml_sycl_device_info info = {};
@@ -195,11 +196,13 @@ static void ggml_check_sycl() try {
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
         g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1);
         g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
+        g_ggml_sycl_prioritize_dmmv = get_sycl_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
         GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
         GGML_LOG_INFO("Running with Environment Variables:\n");
         GGML_LOG_INFO("  GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug);
         GGML_LOG_INFO("  GGML_SYCL_DISABLE_OPT: %d\n", g_ggml_sycl_disable_optimize);
         GGML_LOG_INFO("  GGML_SYCL_DISABLE_GRAPH: %d\n", g_ggml_sycl_disable_graph);
+        GGML_LOG_INFO("  GGML_SYCL_PRIORITIZE_DMMV: %d\n", g_ggml_sycl_prioritize_dmmv);
         GGML_LOG_INFO("Build with Macros:\n");
 #if defined(GGML_SYCL_FORCE_MMQ)
         GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: yes\n");
@@ -2822,12 +2825,45 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     std::exit(1);
 }
 
+enum class mul_mat_algo {
+    DMMV         = 0,
+    MMVQ         = 1,
+    MUL_MAT_SYCL = 2,
+};
+
 inline bool ggml_sycl_supports_mmq(enum ggml_type type) {
     // TODO: accuracy issues in MMQ
     GGML_UNUSED(type);
     return false;
 }
 
+inline bool ggml_sycl_supports_reorder_mul_mat_sycl(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return true;
+        default:
+            return false;
+    }
+}
+
+inline bool ggml_sycl_supports_reorder_dmmv(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return true;
+        default:
+            return false;
+    }
+}
+
+inline bool ggml_sycl_supports_reorder_mmvq(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return true;
+        default:
+            return false;
+    }
+}
+
 static bool ggml_sycl_supports_dmmv(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
@@ -2856,7 +2892,7 @@ static void reorder_qw(char *data_device, const int ncols, const int nrows,
     GGML_ASSERT((size % sizeof(block_q4_0) == 0));
     GGML_ASSERT((offset % sizeof(block_q4_0) == 0));
     int offset_blks = offset / sizeof(block_q4_0);
-    auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;;
+    auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;
     auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks;
 
     stream->parallel_for(
@@ -2884,25 +2920,44 @@ static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
     reorder_qw(data_device, ncols, nrows, size, 0, stream);
 }
 
-/*
-* This function could be called when the OP (mul_mat) function support reorder optimizition.
-*/
-static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1,
-    ggml_tensor * dst) {
-    if (!g_ggml_sycl_disable_optimize && //allow optimize, controlled by $GGML_SYCL_DISABLE_OPT
-        ctx->opt_feature.reorder &&      //allow this device due to good perf, skip the devices with bad perf.
-        dst->op == GGML_OP_MUL_MAT &&    //limit to some supported cases of Q4_0, to do for more cases.
-        src0->type == GGML_TYPE_Q4_0 &&
-        src1->ne[2]==1 && src1->ne[3]==1) {
+static bool should_reorder_tensor(ggml_backend_sycl_context& ctx, const ggml_tensor * dst) {
+    return !g_ggml_sycl_disable_optimize && //allow optimize, controlled by $GGML_SYCL_DISABLE_OPT
+            ctx.opt_feature.reorder &&      //allow this device due to good perf, skip the devices with bad perf.
+            dst->op == GGML_OP_MUL_MAT &&   //limit to some supported cases of Q4_0, to do for more cases.
+            dst->src[1]->ne[2]==1 && dst->src[1]->ne[3]==1;
+}
 
-        ggml_tensor_extra_gpu* extra = (ggml_tensor_extra_gpu*)src0->extra;
-        if (!extra) return; //only happen in CI/UT permute case.
-
-        if (extra->optimized_feature.reorder) return; //skip the tensor which is handled for reorder.
-
-        reorder_qw(src0, ctx->stream());
-        extra->optimized_feature.reorder = true; //used to decode/dequan in next steps.
+static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor * src0, const ggml_tensor * /* src1 */,
+                            ggml_tensor * dst, mul_mat_algo mm_algorithm) {
+    if (!should_reorder_tensor(*ctx, dst)) {
+        return;
     }
+
+    ggml_tensor_extra_gpu * extra = static_cast<ggml_tensor_extra_gpu *>(src0->extra);
+    if (!extra || extra->optimized_feature.reorder) {
+        return;  // Skip permutations and already reordered tensors
+    }
+
+    switch (mm_algorithm) {
+        case mul_mat_algo::DMMV:
+            if (!ggml_sycl_supports_reorder_dmmv(src0->type)) {
+                return;
+            }
+            break;
+        case mul_mat_algo::MMVQ:
+            if (!ggml_sycl_supports_reorder_mmvq(src0->type)) {
+                return;
+            }
+            break;
+        case mul_mat_algo::MUL_MAT_SYCL:
+            if (!ggml_sycl_supports_reorder_mul_mat_sycl(src0->type)) {
+                return;
+            }
+            break;
+    }
+
+    reorder_qw(src0, ctx->stream());
+    extra->optimized_feature.reorder = true;  // Used to decode/dequan in next steps and avoid re-reordering
 }
 
 static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -2911,7 +2966,8 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     int64_t min_compute_capability = INT_MAX;
 
     if (split) {
-        ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *) src0->buffer->buft->context;
+        ggml_backend_sycl_split_buffer_type_context * buft_ctx =
+            (ggml_backend_sycl_split_buffer_type_context *) src0->buffer->buft->context;
         auto & tensor_split = buft_ctx->tensor_split;
         for (int id = 0; id < ggml_sycl_info().device_count; ++id) {
             // skip devices that are not going to do any work:
@@ -2924,7 +2980,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
             }
         }
     } else {
-        min_compute_capability    = ggml_sycl_info().devices[ctx.device].cc;
+        min_compute_capability = ggml_sycl_info().devices[ctx.device].cc;
     }
 
     // check data types and tensor shapes for custom matrix multiplication kernels:
@@ -2946,9 +3002,15 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     use_mul_mat_q = use_mul_mat_q && (src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
 #endif // SYCL_USE_XMX
 
+
     // mmvq path is faster in the CUDA backend.
-    if (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda)
+    if (!g_ggml_sycl_prioritize_dmmv && (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda
+        // Dispatch becomes obscure with the reorder, MMVQ when the reorder optimization
+        // is enabled takes precedence over DMMV, the current if-else implementation
+        // requires disabling DMMV if both conditions are met
+        || (should_reorder_tensor(ctx, dst) && ggml_sycl_supports_reorder_mmvq(src0->type)))) {
         use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
+    }
 
     if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
         // TODO: Refactor and cleanup of mul mat dispatching.
@@ -2967,17 +3029,23 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
         // KQ + KQV multi-batch
         ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
-        opt_for_reorder(&ctx, src0, src1, dst); //the OP function in this branch support reorder.
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false);
-        // save_tensor_txt("1/dst_1.txt", (float*) dst->data, src0->ne[1], sizeof(float), ctx.stream());
+        constexpr bool convert_src1_to_q8_1 = false;
+        opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::DMMV);
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, convert_src1_to_q8_1);
     } else if (use_mul_mat_vec_q) {
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true);
+        constexpr bool convert_src1_to_q8_1 = true;
+        opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::MMVQ);
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, convert_src1_to_q8_1);
     } else if (use_mul_mat_q) {
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, true);
+        constexpr bool convert_src1_to_q8_1 = true;
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, convert_src1_to_q8_1);
     } else {
-        opt_for_reorder(&ctx, src0, src1, dst); //the OP function in this branch support reorder.
-        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false);
+        constexpr bool convert_src1_to_q8_1 = false;
+        // MUL_MAT_SYCL supports reorder
+        opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::MUL_MAT_SYCL);
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, convert_src1_to_q8_1);
     }
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
 }
 
 

ggml/src/ggml-sycl/mmvq.cpp

@@ -1,6 +1,60 @@
 #include "mmvq.hpp"
+
+#include "ggml.h"
+#include "common.hpp"
+#include "quants.hpp"
 #include "vecdotq.hpp"
-#include <cassert>
+
+template <typename reorder_vec_dot_q_sycl>
+static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+                                  const int ncols, const int nrows, const sycl::nd_item<3> & nd_item) {
+    using block_type   = ggml_sycl_reordered::block_q_t<reorder_vec_dot_q_sycl::gtype>;
+    using block_traits = typename block_type::traits;
+
+    const auto sg           = nd_item.get_sub_group();
+    const int  sg_range     = sg.get_group_linear_range();
+    const int  workgroup_id = nd_item.get_group_linear_id();
+    const int  sg_id        = sg.get_group_linear_id();
+    const int  row          = workgroup_id * sg_range + sg_id;
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int     blocks_per_row              = ncols / block_traits::qk;
+    constexpr int blocks_per_subgroup         = ceil_div(block_traits::vdr_mmvq * WARP_SIZE, block_traits::qi);
+    constexpr int block_elements_per_subgroup = block_traits::qi / block_traits::vdr_mmvq;
+
+    static_assert(blocks_per_subgroup > 0);
+    static_assert(block_elements_per_subgroup > 0);
+
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    float partial_sum = 0.0f;
+    for (int i = sg.get_local_linear_id() / block_elements_per_subgroup; i < blocks_per_row; i += blocks_per_subgroup) {
+        const int ibx       = row * blocks_per_row + i;  // x block index
+        // TODO: Generalize offsets, right now only works for quantizations that don't split high and low bits
+        const int bx_offset = block_type::get_block_offset(ibx);
+        const int d_offset  = block_type::get_d_offset(nrows, ncols, ibx);
+
+        // Y block index that aligns with ibx
+        const int iby = i * block_type::block_to_q8_1_ratio();
+
+#pragma unroll
+        for (int elem = 0; elem < block_elements_per_subgroup; elem += WARP_SIZE) {
+            // x block quant index when casting the quants to int
+            const int iqs = elem + block_traits::vdr_mmvq * (sg.get_local_linear_id() % block_elements_per_subgroup);
+
+            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs);
+        }
+    }
+
+    auto sum = sycl::reduce_over_group(nd_item.get_sub_group(), partial_sum, std::plus<>());
+
+    if (sg.leader()) {
+        dst[row] = sum;
+    }
+}
 
 template <int qk, int qi, typename block_q_t, int vdr, vec_dot_q_sycl_t vec_dot_q_sycl>
 static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
@@ -480,26 +534,39 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx,
     }
 }
 
-static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
+static void reorder_mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols,
+                                                    const int nrows, dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK4_0 == 0);
+    const int        block_num_y   = ceil_div(nrows, GGML_SYCL_MMV_Y);
+    constexpr size_t num_subgroups = 16;
+    GGML_ASSERT(block_num_y % num_subgroups == 0);
+
+    const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, (block_num_y * WARP_SIZE));
+    const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
+                         [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0>>(vx, vy, dst, ncols, nrows,
+                                                                                           nd_item);
+                         });
+    });
+}
+
+static void mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols, const int nrows,
                                        dpct::queue_ptr stream) {
     GGML_ASSERT(ncols % QK4_0 == 0);
     const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK4_0, QI4_0, block_q4_0,
-                                      VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        stream->submit([&](sycl::handler & cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                 mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
+                                     vx, vy, dst, ncols, nrows, item_ct1);
+                             });
         });
     }
 }
@@ -916,93 +983,95 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy,
     }
 }
 
-void ggml_sycl_op_mul_mat_vec_q(
-    ggml_backend_sycl_context & ctx,
-    const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
-    const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i,
-    float *dst_dd_i, const int64_t row_low, const int64_t row_high,
-    const int64_t src1_ncols, const int64_t src1_padded_col_size,
-    const dpct::queue_ptr &stream) {
-
+void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1,
+                                ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+                                const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low,
+                                const int64_t row_high, const int64_t src1_ncols, const int64_t src1_padded_col_size,
+                                const dpct::queue_ptr & stream) {
     const int64_t ne10 = src1->ne[0];
     GGML_ASSERT(ne10 % QK8_1 == 0);
 
-    const int64_t ne00 = src0->ne[0];
+    const int64_t ne00     = src0->ne[0];
     const int64_t row_diff = row_high - row_low;
 
     int id;
-    SYCL_CHECK(
-        CHECK_TRY_ERROR(id = get_current_device_id()));
+    SYCL_CHECK(CHECK_TRY_ERROR(id = get_current_device_id()));
     const size_t q8_1_ts = sizeof(block_q8_1);
     const size_t q8_1_bs = QK8_1;
     // the main device has a larger memory buffer to hold the results from all GPUs
     // nrows_dst == nrows of the matrix that the kernel writes into
 
-    for (int i = 0; i < src1_ncols; i++)
-    {
+    for (int i = 0; i < src1_ncols; i++) {
         const size_t src1_ddq_i_offset = i * src1_padded_col_size * q8_1_ts / q8_1_bs;
-        const char* src1_ddq_i_bs = src1_ddq_i + src1_ddq_i_offset;
-        float* dst_dd_i_bs = dst_dd_i + i * dst->ne[0];
+        const char * src1_ddq_i_bs     = src1_ddq_i + src1_ddq_i_offset;
+        float *      dst_dd_i_bs       = dst_dd_i + i * dst->ne[0];
         switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q2_K:
-            mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q3_K:
-            mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_K:
-            mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_K:
-            mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q6_K:
-            mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_IQ1_S:
-            mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_IQ1_M:
-            mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_IQ2_XXS:
-            mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_IQ2_XS:
-            mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_IQ2_S:
-            mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_IQ3_XXS:
-            mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_IQ3_S:
-            mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_IQ4_NL:
-            mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_IQ4_XS:
-            mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
-            break;
-        default:
-            GGML_ABORT("fatal error");
+            case GGML_TYPE_Q4_0:
+                if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
+                    ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                    GGML_SYCL_DEBUG("Calling reorder_mul_mat_vec_q4_0_q8_1_sycl\n");
+                    reorder_mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                } else {
+                    GGML_SYCL_DEBUG("Calling mul_mat_vec_q4_0_q8_1_sycl\n");
+                    mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                }
+                break;
+            case GGML_TYPE_Q4_1:
+                mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_Q5_0:
+                mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_Q5_1:
+                mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_Q8_0:
+                mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_Q2_K:
+                mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_Q3_K:
+                mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_Q4_K:
+                mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_Q5_K:
+                mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_Q6_K:
+                mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_IQ1_S:
+                mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_IQ1_M:
+                mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_IQ2_XXS:
+                mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_IQ2_XS:
+                mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_IQ2_S:
+                mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_IQ3_XXS:
+                mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_IQ3_S:
+                mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_IQ4_NL:
+                mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            case GGML_TYPE_IQ4_XS:
+                mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
+            default:
+                GGML_ABORT("fatal error");
         }
     }
     GGML_UNUSED(src1);

ggml/src/ggml-sycl/quants.hpp

@@ -0,0 +1,61 @@
+//
+// MIT license
+// Copyright (C) 2025 Codeplay Software Ltd.
+// Copyright (C) 2025 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_QUANTS_HPP
+#define GGML_SYCL_QUANTS_HPP
+
+#include "ggml-common.h"
+#include "ggml.h"
+
+namespace ggml_sycl_reordered {
+
+
+// The reordered block moves quants (qs) and  scales(d) to two
+// uniform regions of memory that is contiguous in the same tensor.
+// What this means is that instead of having:
+// [d0, qs0] [d1, qs1] [d2, qs2] ... [dN, qsN]
+// We have:
+// [qs0, qs1, qs2, ..., qsN]  [d0, d1, d2, ..., dN]
+//
+// Notes: out-of-bounds qs will run into d values
+// Aligment relies on the allocated size of qs
+
+template <ggml_type type> struct block_q_t;
+
+
+// qk number of weights / quants in a block
+// qr number of weights in a byte (described as 'before dequantization')
+//    for quantization types that has low and high bits split, qr is calculated with
+//    using the lower bits, e.g for Q6 quants QR6 is 2
+// qi number of 32 bit integers needed to represent all the quants from a block (`qs` field)
+// See ggml-common.h to see how these are calculated
+template <> struct block_q_t<GGML_TYPE_Q4_0> {
+    struct traits {
+        static constexpr uint32_t qk       = QK4_0;
+        static constexpr uint32_t qi       = QI4_0;
+        static constexpr uint32_t qr       = QR4_0;
+        static constexpr uint32_t vdr_mmvq = 2;
+    };
+
+    static constexpr int get_block_offset(const int block_index) { return block_index * (traits::qk / traits::qr); }
+
+    static constexpr int get_d_offset(int nrows, int ncols, const int block_index) {
+        return (ncols / traits::qr * nrows) + block_index * sizeof(ggml_half);
+    }
+
+    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
+};
+
+}  // namespace ggml_sycl_reordered
+
+#endif  // GGML_SYCL_QUANTS_HPP

ggml/src/ggml-sycl/vecdotq.hpp

@@ -1,6 +1,6 @@
 //
 // MIT license
-// Copyright (C) 2024 Intel Corporation
+// Copyright (C) 2025 Intel Corporation
 // SPDX-License-Identifier: MIT
 //
 
@@ -14,8 +14,11 @@
 #define GGML_SYCL_VECDOTQ_HPP
 
 #include "dpct/helper.hpp"
+#include "ggml.h"
+#include "quants.hpp"
 
-typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
+typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1,
+                                  const int & iqs);
 
 static __dpct_inline__ int get_int_from_int8(const int8_t* x8, const int& i32) {
   const uint16_t* x16 =
@@ -252,13 +255,60 @@ vec_dot_q6_K_q8_1_impl_mmvq(const int &vl, const int &vh,
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
+template <ggml_type T> struct reorder_vec_dot_q_sycl {
+    static_assert(T != T, "ggml_type for reorder vecdot not implemented");
+};
+
+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q4_0;
+
+    using q4_0_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q4_0>;
+    using q4_0_traits = typename q4_0_block::traits;
+
+    __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int * v, const int * u, const float & d4, const sycl::half2 & ds8) {
+        int sumi = 0;
+
+#pragma unroll
+        for (size_t i = 0; i < q4_0_traits::vdr_mmvq; ++i) {
+            const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
+            const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
+
+            // SIMD dot product of quantized values
+            sumi = dpct::dp4a(vi0, u[2 * i + 0], sumi);
+            sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi);
+        }
+
+        const sycl::float2 ds8f = ds8.convert<float, sycl::rounding_mode::automatic>();
+
+        // second part effectively subtracts 8 from each quant value
+        return d4 * (sumi * ds8f.x() - (8 * q4_0_traits::vdr_mmvq / q4_0_traits::qi) * ds8f.y());
+    }
+
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
+                     const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+        const uint8_t * bq4_0 = static_cast<const uint8_t *>(vbq) + ibx_offset;
+        const ggml_half d     = *(reinterpret_cast<const ggml_half *>(static_cast<const uint8_t *>(vbq) + d_offset));
+        int             v[q4_0_traits::vdr_mmvq];
+        int             u[2 * q4_0_traits::vdr_mmvq];
+
+#pragma unroll
+
+        for (size_t i = 0; i < q4_0_traits::vdr_mmvq; ++i) {
+            v[i]         = get_int_from_uint8(bq4_0, iqs + i);
+            u[2 * i + 0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+            u[2 * i + 1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + q4_0_traits::qi);
+        }
+
+        return vec_dot_q4_0_q8_1_impl(v, u, d, bq8_1->ds);
+    };
+};
+
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 
 template <int vdr>
-static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int *v, const int *u,
-                                                    const float &d4,
-                                                    const sycl::half2 &ds8) {
+static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int * v, const int * u, const float & d4,
+                                                    const sycl::half2 & ds8) {
     int sumi = 0;
 #pragma unroll
     for (int i = 0; i < vdr; ++i) {
@@ -270,8 +320,7 @@ static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int *v, const int *u,
         sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi);
     }
 
-    const sycl::float2 ds8f =
-        ds8.convert<float, sycl::rounding_mode::automatic>();
+    const sycl::float2 ds8f = ds8.convert<float, sycl::rounding_mode::automatic>();
 
     // second part effectively subtracts 8 from each quant value
     return d4 * (sumi * ds8f.x() - (8 * vdr / QI4_0) * ds8f.y());
@@ -456,13 +505,13 @@ vec_dot_q4_0_q8_1(const void *__restrict__ vbq,
     const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq;
 
     int v[VDR_Q4_0_Q8_1_MMVQ];
-    int u[2*VDR_Q4_0_Q8_1_MMVQ];
+    int u[2 * VDR_Q4_0_Q8_1_MMVQ];
 
 #pragma unroll
     for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) {
-        v[i]     = get_int_from_uint8(bq4_0->qs, iqs + i);
-        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
-        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0);
+        v[i]         = get_int_from_uint8(bq4_0->qs, iqs + i);
+        u[2 * i + 0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+        u[2 * i + 1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0);
     }
 
     return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMVQ>(v, u, bq4_0->d, bq8_1->ds);

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -275,6 +275,7 @@ struct vk_device_struct {
     bool prefer_host_memory;
     bool float_controls_rte_fp16;
     bool subgroup_add;
+    bool subgroup_shuffle;
 
     bool integer_dot_product;
 
@@ -402,12 +403,20 @@ struct vk_device_struct {
     vk_pipeline pipeline_conv2d_dw_cwhn_f32;
 
     // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
+    vk_pipeline pipeline_flash_attn_f32_f16_D64_cm2[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D80_cm2[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D96_cm2[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D112_cm2[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D128_cm2[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D256_cm2[GGML_TYPE_COUNT][2][2][2];
+
     vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
     vk_pipeline pipeline_flash_attn_f32_f16_D80[GGML_TYPE_COUNT][2][2][2];
     vk_pipeline pipeline_flash_attn_f32_f16_D96[GGML_TYPE_COUNT][2][2][2];
     vk_pipeline pipeline_flash_attn_f32_f16_D112[GGML_TYPE_COUNT][2][2][2];
     vk_pipeline pipeline_flash_attn_f32_f16_D128[GGML_TYPE_COUNT][2][2][2];
     vk_pipeline pipeline_flash_attn_f32_f16_D256[GGML_TYPE_COUNT][2][2][2];
+
     vk_pipeline pipeline_flash_attn_split_k_reduce;
 
     std::unordered_map<std::string, vk_pipeline_ref> pipelines;
@@ -1581,13 +1590,29 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
 
 // number of rows/cols for flash attention shader
 static constexpr uint32_t flash_attention_num_small_rows = 32;
-static std::array<uint32_t, 2> fa_rows_cols(uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) {
+static constexpr uint32_t scalar_flash_attention_num_small_rows = 1;
+static constexpr uint32_t scalar_flash_attention_num_large_rows = 8;
+
+static uint32_t get_fa_num_small_rows(bool scalar) {
+    return scalar ? scalar_flash_attention_num_small_rows : flash_attention_num_small_rows;
+}
+
+static std::array<uint32_t, 2> fa_rows_cols(bool scalar, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) {
     GGML_UNUSED(clamp);
 
+    if (scalar) {
+        if (small_rows) {
+            return {scalar_flash_attention_num_small_rows, 64};
+        } else {
+            return {scalar_flash_attention_num_large_rows, 32};
+        }
+    }
+
     // small rows, large cols
     if (small_rows) {
-        return {flash_attention_num_small_rows, 64};
+        return {get_fa_num_small_rows(scalar), 32};
     }
+
     // small cols to reduce register count
     if (ggml_is_quantized(type) || D == 256) {
         return {64, 32};
@@ -1632,7 +1657,7 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
     const uint32_t warps = warptile[0] / warptile[10];
 
     const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size;
-    const uint32_t mmid_row_ids = mul_mat_id ? 3072 * sizeof(uint32_t) : 0;
+    const uint32_t mmid_row_ids = mul_mat_id ? 4096 * sizeof(uint32_t) : 0;
     const uint32_t coopmat_stage = device->coopmat_support ? warptile[7] * warptile[8] / warps * sizeof(float) : 0;
 
     const uint32_t total_size = load_bufs + mmid_row_ids + coopmat_stage + lut_size;
@@ -1882,65 +1907,66 @@ static void ggml_vk_load_shaders(vk_device& device) {
                                       parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size));
     };
 
+    auto const &fa_wg_denoms = [&](bool scalar, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
+        return {fa_rows_cols(scalar, D, clamp, type, small_rows)[0], 1, 1};
+    };
+
+    auto const &fa_spec_constants = [&](bool scalar, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
+        // For large number of rows, 128 invocations seems to work best.
+        // For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we
+        // can't use 256 for D==80.
+        // For scalar, use 128 (arbitrary)
+        uint32_t wg_size = scalar ? 128 : ((small_rows && (D % 32) == 0) ? 256 : 128);
+        auto rows_cols = fa_rows_cols(scalar, D, clamp, type, small_rows);
+
+        // D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it.
+        // D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader.
+        const uint32_t D_lsb = D ^ (D & (D-1));
+        uint32_t D_split = std::min(std::min(device->subgroup_size, 8u), D_lsb / 4);
+
+        // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads
+        GGML_ASSERT((GGML_KQ_MASK_PAD % rows_cols[0]) == 0);
+        return {wg_size, rows_cols[0], rows_cols[1], (D), clamp, D_split};
+    };
+
+#define CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, D) \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,false), fa_spec_constants(SCALAR, D,1,TYPE,false), 1, true);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,false), fa_spec_constants(SCALAR, D,0,TYPE,false), fa_rows_cols(SCALAR,D,0,TYPE,false)[1], true);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,false), fa_spec_constants(SCALAR, D,1,TYPE,false), 1, true);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,false), fa_spec_constants(SCALAR, D,0,TYPE,false), fa_rows_cols(SCALAR,D,0,TYPE,false)[1], true);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,true), fa_spec_constants(SCALAR, D,1,TYPE,true), 1, true);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,true), fa_spec_constants(SCALAR, D,0,TYPE,true), fa_rows_cols(SCALAR,D,0,TYPE,true)[1], true);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,true), fa_spec_constants(SCALAR, D,1,TYPE,true), 1, true);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,true), fa_spec_constants(SCALAR, D,0,TYPE,true), fa_rows_cols(SCALAR,D,0,TYPE,true)[1], true);     \
+
+#define CREATE_FA(TYPE, NAMELC, SCALAR, SUFFIX) \
+        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 64) \
+        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 80) \
+        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 96) \
+        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 112) \
+        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 128) \
+        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 256)
+
+    CREATE_FA(GGML_TYPE_F16, f16, true, )
+    CREATE_FA(GGML_TYPE_Q4_0, q4_0, true, )
+    CREATE_FA(GGML_TYPE_Q8_0, q8_0, true, )
 #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
     if (device->coopmat2) {
-
-        auto const &fa_wg_denoms = [&](uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
-            return {fa_rows_cols(D, clamp, type, small_rows)[0], 1, 1};
-        };
-
-        auto const &fa_spec_constants = [&](uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
-            // For large number of rows, 128 invocations seems to work best.
-            // For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we
-            // can't use 256 for D==80.
-            uint32_t wg_size = (small_rows && (D % 32) == 0) ? 256 : 128;
-            auto rows_cols = fa_rows_cols(D, clamp, type, small_rows);
-            // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads
-            GGML_ASSERT((GGML_KQ_MASK_PAD % rows_cols[0]) == 0);
-            return {wg_size, rows_cols[0], rows_cols[1], (D), clamp};
-        };
-
-#define CREATE_FA2(TYPE, NAMELC, D) \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc"         #NAMELC,           flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,false), fa_spec_constants(D,1,TYPE,false), 1);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC,           flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,false), fa_spec_constants(D,0,TYPE,false), fa_rows_cols(D,0,TYPE,false)[1]);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc"         #NAMELC,           flash_attn_f32_f16_ ## NAMELC ## _cm2_len,         flash_attn_f32_f16_ ## NAMELC ## _cm2_data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,false), fa_spec_constants(D,1,TYPE,false), 1);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC,           flash_attn_f32_f16_ ## NAMELC ## _cm2_len,         flash_attn_f32_f16_ ## NAMELC ## _cm2_data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,false), fa_spec_constants(D,0,TYPE,false), fa_rows_cols(D,0,TYPE,false)[1]);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,true), fa_spec_constants(D,1,TYPE,true), 1);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,true), fa_spec_constants(D,0,TYPE,true), fa_rows_cols(D,0,TYPE,true)[1]);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _cm2_len,         flash_attn_f32_f16_ ## NAMELC ## _cm2_data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,true), fa_spec_constants(D,1,TYPE,true), 1);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _cm2_len,         flash_attn_f32_f16_ ## NAMELC ## _cm2_data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,true), fa_spec_constants(D,0,TYPE,true), fa_rows_cols(D,0,TYPE,true)[1]);     \
-
-#define CREATE_FA(TYPE, NAMELC) \
-        CREATE_FA2(TYPE, NAMELC, 64) \
-        CREATE_FA2(TYPE, NAMELC, 80) \
-        CREATE_FA2(TYPE, NAMELC, 96) \
-        CREATE_FA2(TYPE, NAMELC, 112) \
-        CREATE_FA2(TYPE, NAMELC, 128) \
-        CREATE_FA2(TYPE, NAMELC, 256)
-
-        CREATE_FA(GGML_TYPE_F16, f16)
-        CREATE_FA(GGML_TYPE_Q4_0, q4_0)
-        CREATE_FA(GGML_TYPE_Q4_1, q4_1)
-        CREATE_FA(GGML_TYPE_Q5_0, q5_0)
-        CREATE_FA(GGML_TYPE_Q5_1, q5_1)
-        CREATE_FA(GGML_TYPE_Q8_0, q8_0)
-        // K dequants currently disabled because D dimension is rounded up to 256 and runs inefficiently
-        //CREATE_FA(GGML_TYPE_Q2_K, q2_k)
-        //CREATE_FA(GGML_TYPE_Q3_K, q3_k)
-        //CREATE_FA(GGML_TYPE_Q4_K, q4_k)
-        //CREATE_FA(GGML_TYPE_Q5_K, q5_k)
-        //CREATE_FA(GGML_TYPE_Q6_K, q6_k)
-        //CREATE_FA(GGML_TYPE_IQ1_S, iq1_s)
-        //CREATE_FA(GGML_TYPE_IQ1_M, iq1_m)
-        //CREATE_FA(GGML_TYPE_IQ2_XXS, iq2_xxs)
-        //CREATE_FA(GGML_TYPE_IQ2_XS, iq2_xs)
-        //CREATE_FA(GGML_TYPE_IQ2_S, iq2_s)
-        //CREATE_FA(GGML_TYPE_IQ3_XXS, iq3_xxs)
-        //CREATE_FA(GGML_TYPE_IQ3_S, iq3_s)
-        //CREATE_FA(GGML_TYPE_IQ4_XS, iq4_xs)
-        CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl)
+        CREATE_FA(GGML_TYPE_F16, f16, false, _cm2)
+        CREATE_FA(GGML_TYPE_Q4_0, q4_0, false, _cm2)
+        CREATE_FA(GGML_TYPE_Q4_1, q4_1, false, _cm2)
+        CREATE_FA(GGML_TYPE_Q5_0, q5_0, false, _cm2)
+        CREATE_FA(GGML_TYPE_Q5_1, q5_1, false, _cm2)
+        CREATE_FA(GGML_TYPE_Q8_0, q8_0, false, _cm2)
+        CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl, false, _cm2)
+    }
+#endif
+#undef CREATE_FA2
 #undef CREATE_FA
 
+#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    if (device->coopmat2) {
+
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
         ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
@@ -2837,6 +2863,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->subgroup_add = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);
 
+        device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
+                                   (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle);
+
         const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr;
 
         device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
@@ -5260,7 +5289,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     const uint64_t nei0 = ids->ne[0];
     const uint64_t nei1 = ids->ne[1];
-    GGML_ASSERT(nei0 * nei1 <= 3072);
+    GGML_ASSERT(nei0 * nei1 <= 4096);
 
     const uint32_t nbi1 = ids->nb[1];
     const uint32_t nbi2 = ids->nb[2];
@@ -5709,20 +5738,57 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     assert(q->type == GGML_TYPE_F32);
     assert(k->type == v->type);
 
+    bool scalar = !ctx->device->coopmat2;
+
+    uint32_t gqa_ratio = 1;
+    uint32_t qk_ratio = neq2 / nek2;
+    uint32_t workgroups_x = (uint32_t)neq1;
+    uint32_t workgroups_y = (uint32_t)neq2;
+    uint32_t workgroups_z = (uint32_t)neq3;
+
+    // For scalar FA, we can use the "large" size to accommodate qga.
+    // For coopmat FA, we always use the small size (which is still pretty large for gqa).
+    const uint32_t max_gqa = scalar ? scalar_flash_attention_num_large_rows : get_fa_num_small_rows(false);
+
+    if (N == 1 && qk_ratio > 1 && qk_ratio <= max_gqa &&
+        qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 1) {
+        // grouped query attention - make the N dimension equal to gqa_ratio, reduce
+        // workgroups proportionally in y dimension. The shader will detect gqa_ratio > 1
+        // and change addressing calculations to index Q's dimension 2.
+        gqa_ratio = qk_ratio;
+        N = gqa_ratio;
+        workgroups_y /= N;
+    }
+
     vk_pipeline *pipelines;
     // XXX TODO other backends may be changing accumulator precision to default to f32 soon
-    bool f32acc = dst->op_params[3] == GGML_PREC_F32;
-    bool small_rows = N <= flash_attention_num_small_rows;
-    switch (D) {
-    case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64[k->type][f32acc][small_rows][0]; break;
-    case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80[k->type][f32acc][small_rows][0]; break;
-    case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96[k->type][f32acc][small_rows][0]; break;
-    case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112[k->type][f32acc][small_rows][0]; break;
-    case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128[k->type][f32acc][small_rows][0]; break;
-    case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256[k->type][f32acc][small_rows][0]; break;
-    default:
-        assert(!"unsupported D value");
-        return;
+    bool f32acc = scalar || dst->op_params[3] == GGML_PREC_F32;
+    bool small_rows = N <= get_fa_num_small_rows(scalar);
+
+    if (scalar) {
+        switch (D) {
+        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64[k->type][f32acc][small_rows][0]; break;
+        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80[k->type][f32acc][small_rows][0]; break;
+        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96[k->type][f32acc][small_rows][0]; break;
+        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112[k->type][f32acc][small_rows][0]; break;
+        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128[k->type][f32acc][small_rows][0]; break;
+        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256[k->type][f32acc][small_rows][0]; break;
+        default:
+            GGML_ASSERT(!"unsupported D value");
+            return;
+        }
+    } else {
+        switch (D) {
+        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm2[k->type][f32acc][small_rows][0]; break;
+        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm2[k->type][f32acc][small_rows][0]; break;
+        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm2[k->type][f32acc][small_rows][0]; break;
+        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm2[k->type][f32acc][small_rows][0]; break;
+        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm2[k->type][f32acc][small_rows][0]; break;
+        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm2[k->type][f32acc][small_rows][0]; break;
+        default:
+            GGML_ASSERT(!"unsupported D value");
+            return;
+        }
     }
     assert(pipelines);
 
@@ -5740,27 +5806,14 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     vk_pipeline pipeline = pipelines[aligned];
     assert(pipeline);
 
-    uint32_t gqa_ratio = 1;
-    uint32_t qk_ratio = neq2 / nek2;
-    uint32_t workgroups_x = (uint32_t)neq1;
-    uint32_t workgroups_y = (uint32_t)neq2;
-    uint32_t workgroups_z = (uint32_t)neq3;
-
-    if (N == 1 && qk_ratio > 1 && gqa_ratio <= flash_attention_num_small_rows &&
-        qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 1) {
-        // grouped query attention - make the N dimension equal to gqa_ratio, reduce
-        // workgroups proportionally in y dimension. The shader will detect gqa_ratio > 1
-        // and change addressing calculations to index Q's dimension 2.
-        gqa_ratio = qk_ratio;
-        N = gqa_ratio;
-        workgroups_y /= N;
-    }
-
     uint32_t split_kv = KV;
     uint32_t split_k = 1;
 
+    // Use a placeholder core count if one isn't available. split_k is a big help for perf.
+    const uint32_t shader_core_count = ctx->device->shader_core_count ? ctx->device->shader_core_count : 16;
+
     // Try to use split_k when KV is large enough to be worth the overhead
-    if (workgroups_x == 1 && ctx->device->shader_core_count > 0 && KV >= 512) {
+    if (workgroups_x == 1 && shader_core_count > 0 && KV >= 512) {
         // Try to run two workgroups per SM.
         split_k = ctx->device->shader_core_count * 2 / workgroups_y;
         if (split_k > 1) {
@@ -9530,9 +9583,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-                if (!ggml_vk_get_device(ctx->device)->coopmat2) {
-                    return false;
-                }
+                auto device = ggml_vk_get_device(ctx->device);
+                bool coopmat2 = device->coopmat2;
                 switch (op->src[0]->ne[0]) {
                 case 64:
                 case 80:
@@ -9540,7 +9592,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 case 112:
                 case 128:
                 case 256:
-                case 575: // DeepSeek MLA
                     break;
                 default:
                     return false;
@@ -9566,10 +9617,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 switch (op->src[1]->type) {
                 case GGML_TYPE_F16:
                 case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q8_0:
+                    // supported in scalar and coopmat2 paths
+                    break;
                 case GGML_TYPE_Q4_1:
                 case GGML_TYPE_Q5_0:
                 case GGML_TYPE_Q5_1:
-                case GGML_TYPE_Q8_0:
                 // K dequants currently disabled because D dimension is rounded up to 256 and runs inefficiently
                 //case GGML_TYPE_Q2_K:
                 //case GGML_TYPE_Q3_K:
@@ -9585,10 +9638,18 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 //case GGML_TYPE_IQ3_S:
                 //case GGML_TYPE_IQ4_XS:
                 case GGML_TYPE_IQ4_NL:
+                    // currently supported only in coopmat2 path
+                    if (!coopmat2) {
+                        return false;
+                    }
                     break;
                 default:
                     return false;
                 }
+                if (!coopmat2 && !device->subgroup_shuffle) {
+                    // scalar FA uses subgroupShuffle
+                    return false;
+                }
                 return true;
             }
         case GGML_OP_GET_ROWS:

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

@@ -0,0 +1,483 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#extension GL_KHR_shader_subgroup_shuffle : enable
+
+#include "types.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 1) const uint32_t Br = 1;
+layout (constant_id = 2) const uint32_t Bc = 32;
+layout (constant_id = 3) const uint32_t D = 32;
+
+layout (constant_id = 5) const uint32_t D_split = 16;
+const uint32_t D_per_thread = D / D_split;
+
+const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split;
+const uint32_t cols_per_thread = Bc / cols_per_iter;
+
+layout (push_constant) uniform parameter {
+    uint32_t N;
+    uint32_t KV;
+
+    uint32_t ne1;
+    uint32_t ne2;
+    uint32_t ne3;
+
+    uint32_t neq2;
+    uint32_t neq3;
+    uint32_t nek2;
+    uint32_t nek3;
+    uint32_t nev2;
+    uint32_t nev3;
+    uint32_t nem1;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+    uint32_t nb21;
+    uint32_t nb22;
+    uint32_t nb23;
+    uint32_t nb31;
+
+    float scale;
+    float max_bias;
+    float logit_softcap;
+
+    uint32_t mask;
+    uint32_t n_head_log2;
+    float m0;
+    float m1;
+
+    uint32_t gqa_ratio;
+    uint32_t split_kv;
+    uint32_t k_num;
+} p;
+
+layout (binding = 0) readonly buffer Q {float data_q[];};
+layout (binding = 0) readonly buffer QV4 {vec4 data_qv4[];};
+layout (binding = 1) readonly buffer K {float16_t data_k[];};
+layout (binding = 1) readonly buffer KV4 {f16vec4 data_kv4[];};
+layout (binding = 2) readonly buffer V {float16_t data_v[];};
+layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];};
+layout (binding = 3) readonly buffer M {float16_t data_m[];};
+layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
+
+#if defined(A_TYPE_PACKED16)
+#define BINDING_IDX_K 0
+#define BINDING_IDX_V 1
+layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2];
+#endif
+
+#if defined(DATA_A_Q4_0)
+#define BLOCK_BYTE_SIZE 18
+
+vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+    uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
+    uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
+    uint shift = (iqs & 0x10) >> 2;
+    vui_lo >>= shift;
+    vui_hi >>= shift;
+
+    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+#define BLOCK_BYTE_SIZE 34
+vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+    const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
+    const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
+
+    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+}
+#endif
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+// Store the output when doing grouped query attention.
+// Rows index by Q's dimension 2, and the first N rows are valid.
+D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
+{
+    uint32_t offset = (iq2 + r) * D + c;
+    data_o[o_offset + offset] = D_TYPE(elem);
+    return elem;
+}
+
+// Store column zero. This is used to save per-row m and L values for split_k.
+ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
+{
+    if (r < N && c == 0) {
+        uint32_t offset = iq2 + r;
+        data_o[o_offset + offset] = D_TYPE(elem);
+    }
+    return elem;
+}
+
+// Load the slope matrix, indexed by Q's dimension 2.
+ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
+{
+    const uint32_t h = iq2 + (r % p.gqa_ratio);
+
+    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
+    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
+
+    return ACC_TYPE(pow(base, ACC_TYPE(exph)));
+}
+
+shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
+shared vec4 tmpshv4[gl_WorkGroupSize.x];
+
+shared float masksh[Bc][Br];
+shared vec4 Qf[Br][D / 4];
+
+void main() {
+#ifdef NEEDS_INIT_IQ_SHMEM
+    init_iq_shmem(gl_WorkGroupSize);
+#endif
+
+    const uint32_t tid = gl_LocalInvocationIndex;
+    const uint32_t N = p.N;
+    const uint32_t KV = p.KV;
+
+    const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
+    const uint32_t col_tid = gl_LocalInvocationIndex / D_split;
+
+    uint32_t i = gl_WorkGroupID.x;
+    uint32_t split_k_index = 0;
+
+    if (p.k_num > 1) {
+        i = 0;
+        split_k_index = gl_WorkGroupID.x;
+    }
+
+    const uint32_t Tr = CEIL_DIV(N, Br);
+
+    const uint32_t start_j = split_k_index * p.split_kv / Bc;
+    const uint32_t end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc);
+
+    // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y.
+    // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2.
+    const uint32_t iq2 = gl_WorkGroupID.y * p.gqa_ratio;
+    const uint32_t iq3 = gl_WorkGroupID.z;
+
+    // broadcast factors
+    const uint32_t rk2 = p.neq2/p.nek2;
+    const uint32_t rk3 = p.neq3/p.nek3;
+
+    const uint32_t rv2 = p.neq2/p.nev2;
+    const uint32_t rv3 = p.neq3/p.nev3;
+
+    // k indices
+    const uint32_t ik3 = iq3 / rk3;
+    const uint32_t ik2 = iq2 / rk2;
+
+    // v indices
+    const uint32_t iv3 = iq3 / rv3;
+    const uint32_t iv2 = iq2 / rv2;
+
+    // nb?1 are already divided by the type size and are in units of elements.
+    // When using grouped query attention, Q is indexed by iq2, so the stride
+    // should be nb02 (which is in bytes).
+    uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
+    uint32_t k_stride = p.nb11;
+    uint32_t v_stride = p.nb21;
+    // When using grouped query attention, all rows use the same mask (stride 0).
+    // "p.gqa_ratio >> 16" is just a roundabout way of writing zero
+    // that prevents the compiler from folding the "&" through the select
+    // and breaking the alignment detection.
+    uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
+
+    uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
+
+    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
+        uint32_t d = (idx + tid) % (D / 4);
+        uint32_t r = (idx + tid) / (D / 4);
+        if (r < Br && d < D / 4 &&
+            i * Br + r < N) {
+            Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale;
+        }
+    }
+    barrier();
+
+    vec4 Of[Br][D_per_thread / 4];
+    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            Of[r][d] = vec4(0.0);
+        }
+    }
+
+    float Lf[Br], Mf[Br];
+
+    // Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M.
+    const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF);
+
+    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+        Lf[r] = 0;
+        Mf[r] = NEG_FLT_MAX_OVER_2;
+    }
+
+    float slope[Br];
+    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+        slope[r] = 1.0;
+    }
+
+    // ALiBi
+    if (p.max_bias > 0.0f) {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2);
+        }
+    }
+
+#if BLOCK_SIZE > 1
+    uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / BLOCK_BYTE_SIZE;
+    uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / BLOCK_BYTE_SIZE;
+#else
+    uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
+    uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
+#endif
+
+    [[dont_unroll]]
+    for (uint32_t j = start_j; j < end_j; ++j) {
+
+        float Sf[Br][cols_per_thread];
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                Sf[r][c] = 0.0;
+            }
+        }
+
+
+        [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+#if BLOCK_SIZE > 1
+                uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
+                uint ib = coord / BLOCK_SIZE;
+                uint iqs = (coord % BLOCK_SIZE);
+                vec4 K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
+#else
+                vec4 K_Tf = vec4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
+#endif
+                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+                    Sf[r][c] += dot(Qf[r][d * D_split + d_tid], K_Tf);
+                }
+            }
+        }
+
+        [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+            // Compute sum across the D_split
+            [[unroll]] for (uint s = D_split / 2; s > 0; s >>= 1) {
+                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+                    Sf[r][c] += subgroupShuffleXor(Sf[r][c], s);
+                }
+            }
+        }
+
+        if (p.logit_softcap != 0.0f) {
+            [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+                [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                    Sf[r][c] = p.logit_softcap * tanh(Sf[r][c]);
+                }
+            }
+        }
+
+        if (p.mask != 0) {
+
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
+                uint32_t c = (idx + tid) % Bc;
+                uint32_t r = (idx + tid) / Bc;
+                if (idx + tid < Bc * Br) {
+                    masksh[c][r] = float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]);
+                }
+            }
+            barrier();
+
+            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+                    float mvf = masksh[c * cols_per_iter + col_tid][r];
+
+                    Sf[r][c] += slope[r]*mvf;
+                }
+            }
+            barrier();
+        }
+
+        float rowmaxf[Br], Pf[Br][cols_per_thread], rowsumf[Br], eMf[Br], Moldf[Br];
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            rowmaxf[r] = Sf[r][0];
+            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                rowmaxf[r] = max(rowmaxf[r], Sf[r][c]);
+            }
+            Moldf[r] = Mf[r];
+
+            // M = max(rowmax, Mold)
+            // P = e^(S - M)
+            // eM = e^(Mold - M)
+            Mf[r] = max(rowmaxf[r], Moldf[r]);
+            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                Pf[r][c] = exp(Sf[r][c] - Mf[r]);
+            }
+            eMf[r] = exp(Moldf[r] - Mf[r]);
+
+            // Compute sum across row of P
+            rowsumf[r] = 0.0;
+            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                rowsumf[r] += Pf[r][c];
+            }
+
+            Lf[r] = eMf[r]*Lf[r] + rowsumf[r];
+        }
+
+        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+                Of[r][d] = eMf[r] * Of[r][d];
+            }
+        }
+
+        [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+#if BLOCK_SIZE > 1
+                uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
+                uint ib = coord / BLOCK_SIZE;
+                uint iqs = (coord % BLOCK_SIZE);
+                vec4 Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
+#else
+                vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
+#endif
+                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+                    Of[r][d] += Pf[r][c] * Vf;
+                }
+            }
+        }
+
+        barrier();
+    }
+
+    // reduce across threads
+
+    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+        float rowmaxf, eMf;
+
+        tmpsh[tid] = Mf[r];
+        // Compute max across the row
+        barrier();
+        [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
+            if (tid < s) {
+                tmpsh[tid] = max(tmpsh[tid], tmpsh[tid + s]);
+            }
+            barrier();
+        }
+        rowmaxf = tmpsh[d_tid];
+        barrier();
+
+        float Moldf = Mf[r];
+
+        // M = max(rowmax, Mold)
+        // eM = e^(Mold - M)
+        Mf[r] = max(rowmaxf, Moldf);
+        eMf = exp(Moldf - Mf[r]);
+
+        Lf[r] = eMf*Lf[r];
+
+        tmpsh[tid] = Lf[r];
+
+        // Compute sum across the row
+        barrier();
+        [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
+            if (tid < s) {
+                tmpsh[tid] = tmpsh[tid] + tmpsh[tid + s];
+            }
+            barrier();
+        }
+        Lf[r] = tmpsh[d_tid];
+        barrier();
+
+        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+
+            Of[r][d] = eMf * Of[r][d];
+            tmpshv4[tid] = Of[r][d];
+
+            barrier();
+            [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
+                if (tid < s) {
+                    Of[r][d] += tmpshv4[tid + s];
+                    tmpshv4[tid] = Of[r][d];
+                }
+                barrier();
+            }
+            Of[r][d] = tmpshv4[d_tid];
+            barrier();
+        }
+    }
+
+
+    // If there is split_k, then the split_k resolve shader does the final
+    // division by L. Store the intermediate O value and per-row m and L values.
+    if (p.k_num > 1) {
+        uint32_t o_offset = D * p.ne1 * split_k_index;
+
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            if (r < N) {
+                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
+                        perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
+                    }
+                }
+            }
+        }
+
+        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            if (r < N) {
+                perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
+                perElemOpStoreCol0(r, 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
+            }
+        }
+
+        return;
+    }
+
+    float Lfrcp[Br];
+    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+        Lfrcp[r] = 1.0 / Lf[r];
+    }
+
+    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            Of[r][d] *= Lfrcp[r];
+        }
+    }
+
+    uint32_t o_offset = iq3*p.ne2*p.ne1;
+
+    if (p.gqa_ratio > 1) {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            if (r < N) {
+                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
+                        perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
+                    }
+                }
+            }
+        }
+    } else {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            if (i * Br + r < N) {
+                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
+                        data_o[o_offset + iq2 * D + (i * Br + r) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                    }
+                }
+            }
+        }
+    }
+}

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

@@ -103,7 +103,7 @@ shared FLOAT_TYPE buf_a[BM * SHMEM_STRIDE];
 shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE];
 
 #ifdef MUL_MAT_ID
-shared u16vec2 row_ids[3072];
+shared u16vec2 row_ids[4096];
 #endif // MUL_MAT_ID
 
 #define NUM_WARPS (BLOCK_SIZE / WARP)

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -92,7 +92,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
 
-shared u16vec4 row_ids[3072];
+shared u16vec4 row_ids[4096];
 
 layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufB {
    B_TYPE b[];

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

@@ -101,7 +101,7 @@ shared FLOAT_TYPE_VEC2 buf_b_ds[BN];
 #define LOAD_VEC_B 4
 
 #ifdef MUL_MAT_ID
-shared u16vec2 row_ids[3072];
+shared u16vec2 row_ids[4096];
 #endif // MUL_MAT_ID
 
 #define NUM_WARPS (BLOCK_SIZE / WARP)

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -421,7 +421,6 @@ void process_shaders() {
 #endif
     }
 
-#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
     // flash attention
     for (const auto& f16acc : {false, true}) {
         std::string acctype = f16acc ? "float16_t" : "float";
@@ -432,6 +431,7 @@ void process_shaders() {
             }
             if (tname == "bf16") continue;
 
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
             if (tname == "f16") {
                 string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
                     merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, true, f16acc);
@@ -440,9 +440,17 @@ void process_shaders() {
                 string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
                     merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc);
             }
+#endif
+            if (tname == "f16") {
+                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
+                    merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, false, f16acc);
+            } else if (tname == "q4_0" || tname == "q8_0") {
+                std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
+                    merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, false, f16acc);
+            }
         }
     }
-#endif
 
     for (const auto& tname : type_names) {
         // mul mat vec

ggml/src/ggml.c

@@ -2732,11 +2732,11 @@ void ggml_mul_mat_set_prec(
     c = ggml_mul_mat_id(ctx, as, b, ids);
 
     as  -> [cols, rows, n_expert]
-    ids -> [n_experts_used, n_tokens] (i32)
     b   -> [cols, n_expert_used, n_tokens]
+    ids -> [n_expert_used, n_tokens] (i32)
     c   -> [rows, n_expert_used, n_tokens]
 
-    in b, n_experts_used can be broadcasted to match the n_expert_used of ids
+    in b, n_expert_used can be broadcasted to match the n_expert_used of ids
 
     c ~= as[:,:,i] @ b[:,i%r,t], i = ids[e,t] for all e,t in ids
 */

gguf-py/gguf/constants.py

@@ -483,7 +483,9 @@ class MODEL_TENSOR(IntEnum):
     V_ENC_EMBD_PATCH     = auto()
     V_ENC_EMBD_POS       = auto()
     V_ENC_ATTN_Q         = auto()
+    V_ENC_ATTN_Q_NORM    = auto()
     V_ENC_ATTN_K         = auto()
+    V_ENC_ATTN_K_NORM    = auto()
     V_ENC_ATTN_V         = auto()
     V_ENC_INPUT_NORM     = auto()
     V_ENC_OUTPUT         = auto()
@@ -491,6 +493,8 @@ class MODEL_TENSOR(IntEnum):
     V_ENC_FFN_UP         = auto()
     V_ENC_FFN_GATE       = auto()
     V_ENC_FFN_DOWN       = auto()
+    V_LAYER_SCALE_1      = auto()
+    V_LAYER_SCALE_2      = auto()
     V_PRE_NORM           = auto()
     V_POST_NORM          = auto()
     V_MM_INP_NORM        = auto()
@@ -740,7 +744,9 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_ENC_EMBD_PATCH:          "v.patch_embd",
     MODEL_TENSOR.V_ENC_EMBD_POS:            "v.position_embd",
     MODEL_TENSOR.V_ENC_ATTN_Q:              "v.blk.{bid}.attn_q",
+    MODEL_TENSOR.V_ENC_ATTN_Q_NORM:         "v.blk.{bid}.attn_q_norm",
     MODEL_TENSOR.V_ENC_ATTN_K:              "v.blk.{bid}.attn_k",
+    MODEL_TENSOR.V_ENC_ATTN_K_NORM:         "v.blk.{bid}.attn_k_norm",
     MODEL_TENSOR.V_ENC_ATTN_V:              "v.blk.{bid}.attn_v",
     MODEL_TENSOR.V_ENC_INPUT_NORM:          "v.blk.{bid}.ln1",
     MODEL_TENSOR.V_ENC_OUTPUT:              "v.blk.{bid}.attn_out",
@@ -748,6 +754,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_ENC_FFN_UP:              "v.blk.{bid}.ffn_up",
     MODEL_TENSOR.V_ENC_FFN_GATE:            "v.blk.{bid}.ffn_gate",
     MODEL_TENSOR.V_ENC_FFN_DOWN:            "v.blk.{bid}.ffn_down",
+    MODEL_TENSOR.V_LAYER_SCALE_1:           "v.blk.{bid}.ls1",
+    MODEL_TENSOR.V_LAYER_SCALE_2:           "v.blk.{bid}.ls2",
     MODEL_TENSOR.V_PRE_NORM:                "v.pre_ln",
     MODEL_TENSOR.V_POST_NORM:               "v.post_ln",
     MODEL_TENSOR.V_MM_INP_PROJ:             "mm.input_projection",
@@ -778,7 +786,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_ENC_EMBD_PATCH,
         MODEL_TENSOR.V_ENC_EMBD_POS,
         MODEL_TENSOR.V_ENC_ATTN_Q,
+        MODEL_TENSOR.V_ENC_ATTN_Q_NORM,
         MODEL_TENSOR.V_ENC_ATTN_K,
+        MODEL_TENSOR.V_ENC_ATTN_K_NORM,
         MODEL_TENSOR.V_ENC_ATTN_V,
         MODEL_TENSOR.V_ENC_INPUT_NORM,
         MODEL_TENSOR.V_ENC_OUTPUT,
@@ -786,6 +796,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_ENC_FFN_UP,
         MODEL_TENSOR.V_ENC_FFN_GATE,
         MODEL_TENSOR.V_ENC_FFN_DOWN,
+        MODEL_TENSOR.V_LAYER_SCALE_1,
+        MODEL_TENSOR.V_LAYER_SCALE_2,
         MODEL_TENSOR.V_PRE_NORM,
         MODEL_TENSOR.V_POST_NORM,
         MODEL_TENSOR.V_MM_INP_PROJ,
@@ -2167,6 +2179,7 @@ class VisionProjectorType:
     PIXTRAL = "pixtral"
     QWEN2VL = "qwen2vl_merger"
     QWEN25VL = "qwen2.5vl_merger"
+    INTERNVL = "internvl"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/tensor_mapping.py

@@ -905,6 +905,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_MMPROJ_MLP: (
             "model.mm_projector.mlp.mlp.{bid}",
+            "mlp1.{bid}", # InternVL
         ),
 
         MODEL_TENSOR.V_MMPROJ_PEG: (
@@ -937,6 +938,10 @@ class TensorNameMap:
             "visual.blocks.{bid}.attn.q", # qwen2vl, generated
         ),
 
+        MODEL_TENSOR.V_ENC_ATTN_Q_NORM: (
+            "vision_tower.vision_model.encoder.layers.{bid}.attn.q_norm", # InternVL
+        ),
+
         MODEL_TENSOR.V_ENC_ATTN_K: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.k_proj",
             "vpm.encoder.layers.{bid}.self_attn.k_proj",
@@ -945,6 +950,10 @@ class TensorNameMap:
             "visual.blocks.{bid}.attn.k", # qwen2vl, generated
         ),
 
+        MODEL_TENSOR.V_ENC_ATTN_K_NORM: (
+            "vision_tower.vision_model.encoder.layers.{bid}.attn.k_norm", # InternVL
+        ),
+
         MODEL_TENSOR.V_ENC_ATTN_V: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.v_proj",
             "vpm.encoder.layers.{bid}.self_attn.v_proj",
@@ -955,6 +964,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_ENC_INPUT_NORM: (
             "vision_tower.vision_model.encoder.layers.{bid}.layer_norm1",
+            "vision_tower.vision_model.encoder.layers.{bid}.norm1", # InternVL
             "vpm.encoder.layers.{bid}.layer_norm1",
             "model.vision_model.encoder.layers.{bid}.layer_norm1", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention_norm", # pixtral
@@ -963,6 +973,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_ENC_OUTPUT: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.out_proj",
+            "vision_tower.vision_model.encoder.layers.{bid}.attn.proj", # InternVL
             "vpm.encoder.layers.{bid}.self_attn.out_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral
@@ -971,6 +982,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_ENC_OUTPUT_NORM: (
             "vision_tower.vision_model.encoder.layers.{bid}.layer_norm2",
+            "vision_tower.vision_model.encoder.layers.{bid}.norm2", # InternVL
             "vpm.encoder.layers.{bid}.layer_norm2",
             "model.vision_model.encoder.layers.{bid}.layer_norm2", # SmolVLM
             "vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral
@@ -1000,6 +1012,14 @@ class TensorNameMap:
             "visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl
         ),
 
+        MODEL_TENSOR.V_LAYER_SCALE_1: (
+            "vision_tower.vision_model.encoder.layers.{bid}.ls1", # InternVL
+        ),
+
+        MODEL_TENSOR.V_LAYER_SCALE_2: (
+            "vision_tower.vision_model.encoder.layers.{bid}.ls2", # InternVL
+        ),
+
         MODEL_TENSOR.V_PRE_NORM: (
             "vision_tower.vision_model.pre_layrnorm",
             "vision_tower.ln_pre", # pixtral

include/llama.h

@@ -112,6 +112,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
         LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
         LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
+        LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
     };
 
     enum llama_rope_type {
@@ -362,6 +363,7 @@ extern "C" {
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
         bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
         bool no_perf;     // whether to measure performance timings
+        bool op_offload;  // whether to offload host tensor operations to device
     };
 
     // model quantization parameters

src/llama-adapter.cpp

@@ -253,6 +253,9 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
     std::vector<ggml_backend_buffer_type_t> buft_extra;
     {
         auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (!cpu_dev) {
+            throw std::runtime_error(format("%s: no CPU backend found", __func__));
+        }
         auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
 
         auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
@@ -291,6 +294,9 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
                 LLAMA_LOG_WARN("%s: lora for '%s' cannot use buft '%s', fallback to CPU\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
 
                 auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+                if (!cpu_dev) {
+                    throw std::runtime_error(format("%s: no CPU backend found", __func__));
+                }
                 buft = ggml_backend_dev_buffer_type(cpu_dev);
 
                 break;

src/llama-chat.cpp

@@ -35,6 +35,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "mistral-v3",        LLM_CHAT_TEMPLATE_MISTRAL_V3        },
     { "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN },
     { "mistral-v7",        LLM_CHAT_TEMPLATE_MISTRAL_V7        },
+    { "mistral-v7-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN },
     { "phi3",              LLM_CHAT_TEMPLATE_PHI_3             },
     { "phi4",              LLM_CHAT_TEMPLATE_PHI_4             },
     { "falcon3",           LLM_CHAT_TEMPLATE_FALCON_3          },
@@ -202,19 +203,20 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|im_start|>assistant\n";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7 || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN) {
         // Official mistral 'v7' template
         // See: https://huggingface.co/mistralai/Mistral-Large-Instruct-2411#basic-instruct-template-v7
+        //      https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503#basic-instruct-template-v7-tekken
+        const char * trailing_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7 ? " " : "";
         for (auto message : chat) {
             std::string role(message->role);
             std::string content(message->content);
             if (role == "system") {
-                ss << "[SYSTEM_PROMPT] " << content << "[/SYSTEM_PROMPT]";
+                ss << "[SYSTEM_PROMPT]" << trailing_space << content << "[/SYSTEM_PROMPT]";
             } else if (role == "user") {
-                ss << "[INST] " << content << "[/INST]";
-            }
-            else {
-                ss << " " << content << "</s>";
+                ss << "[INST]" << trailing_space << content << "[/INST]";
+            } else {
+                ss << trailing_space << content << "</s>";
             }
         }
     } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1

src/llama-chat.h

@@ -14,6 +14,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MISTRAL_V3,
     LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,
     LLM_CHAT_TEMPLATE_MISTRAL_V7,
+    LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN,
     LLM_CHAT_TEMPLATE_PHI_3,
     LLM_CHAT_TEMPLATE_PHI_4,
     LLM_CHAT_TEMPLATE_FALCON_3,

src/llama-context.cpp

@@ -93,6 +93,7 @@ llama_context::llama_context(
     }
 
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
+    cparams.op_offload = params.op_offload;
 
     const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
 
@@ -243,7 +244,7 @@ llama_context::llama_context(
             }
         }
 
-        sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel));
+        sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel, cparams.op_offload));
 
         if (pipeline_parallel) {
             LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
@@ -1871,6 +1872,7 @@ llama_context_params llama_context_default_params() {
         /*.offload_kqv                 =*/ true,
         /*.flash_attn                  =*/ false,
         /*.no_perf                     =*/ true,
+        /*.op_offload                  =*/ true,
     };
 
     return result;

src/llama-cparams.h

@@ -30,6 +30,7 @@ struct llama_cparams {
     bool flash_attn;
     bool no_perf;
     bool warmup;
+    bool op_offload;
 
     enum llama_pooling_type pooling_type;
 

src/llama-graph.cpp

@@ -1227,8 +1227,19 @@ ggml_tensor * llm_graph_context::build_attn_mha(
         ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
 
         if (v_mla) {
+#if 0
+            // v_mla can be applied as a matrix-vector multiplication with broadcasting across dimension 3 == n_tokens.
+            // However, the code is optimized for dimensions 0 and 1 being large, so this is ineffient.
             cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
             cur = ggml_mul_mat(ctx0, v_mla, cur);
+#else
+            // It's preferable to do the calculation as a matrix-matrix multiplication with n_tokens in dimension 1.
+            // The permutations are noops and only change how the tensor data is interpreted.
+            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+            cur = ggml_mul_mat(ctx0, v_mla, cur);
+            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+            cur = ggml_cont(ctx0, cur); // Needed because ggml_reshape_2d expects contiguous inputs.
+#endif
         }
 
         cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);

src/llama-model-loader.cpp

@@ -823,6 +823,10 @@ void llama_model_loader::init_mappings(bool prefetch, llama_mlocks * mlock_mmaps
         mmaps_used.reserve(files.size());
         for (const auto & file : files) {
             auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
+            if (!reg) {
+                throw std::runtime_error(format("%s: no CPU backend found", __func__));
+            }
+
             auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
             std::unique_ptr<llama_mmap> mapping = std::make_unique<llama_mmap>(file.get(), prefetch ? -1 : 0, is_numa_fn());
             mmaps_used.emplace_back(mapping->size(), 0);

src/llama-model.cpp

@@ -299,6 +299,10 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
     // add extra buffer types, only if no GPU device is present
     // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094
     auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (cpu_dev == nullptr) {
+        throw std::runtime_error(format("%s: no CPU backend found", __func__));
+    }
+
     auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
     auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
         ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
@@ -1484,6 +1488,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     }
 
     ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (cpu_dev == nullptr) {
+        throw std::runtime_error(format("%s: no CPU backend found", __func__));
+    }
     const int i_gpu_start = std::max((int) hparams.n_layer - n_gpu_layers, (int) 0);
     const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, (int)n_layer + 1);
     auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev {
@@ -1672,6 +1679,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             auto * buft_dev = ggml_backend_buft_get_device(buft);
             if (ml.use_mmap && buft_dev && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {
                 auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+                if (!cpu_dev) {
+                    throw std::runtime_error("no CPU backend found");
+                }
                 buft = ggml_backend_dev_buffer_type(cpu_dev);
             }
 
@@ -4122,6 +4132,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         if (!dev) {
             // FIXME: workaround for CPU backend buft having a NULL device
             dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+            if (!dev) {
+                throw std::runtime_error(format("%s: no CPU backend found", __func__));
+            }
         }
         ggml_backend_dev_props props;
         ggml_backend_dev_get_props(dev, &props);
@@ -13387,6 +13400,14 @@ const char * llama_model_chat_template(const llama_model * model, const char * n
         : LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE);
     const auto & it = model->gguf_kv.find(key);
     if (it == model->gguf_kv.end()) {
+        // one-off fix for very popular models (so we are not flooded with issues)
+        // do not extend this list unless absolutely necessary
+        // Mistral-Small-2503 does not have built-in chat template
+        llama_vocab_pre_type pre_type = model->vocab.get_pre_type();
+        if (pre_type == LLAMA_VOCAB_PRE_TYPE_TEKKEN && model->layers.size() == 40) {
+            return "mistral-v7-tekken";
+        }
+
         return nullptr;
     }
 

src/llama-vocab.cpp

@@ -415,6 +415,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "'(?:[sSdDmMtT]|[lL][lL]|[vV][eE]|[rR][eE])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_SEED_CODER:
+                regex_exprs = {
+                    // original regex from tokenizer.json
+                    // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1}| ?[^\\s\\p{L}\\p{N}\r\n]+|\\s*[\r\n]+|\\s+(?!\\S)|\\s+"
+                    "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1}| ?[^\\s\\p{L}\\p{N}\\r\\n]+|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1634,6 +1641,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "bailingmoe") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "seed-coder") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_SEED_CODER;
+                clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }

tests/test-opt.cpp

@@ -853,7 +853,7 @@ int main(void) {
         backends_modded.insert(backends_modded.end(), backends.begin(), backends.end());
 
         ggml_backend_sched_t backend_sched = ggml_backend_sched_new(
-            backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false);
+            backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false, true);
 
         printf("Backend %zu/%zu: %s\n", i + 1, dev_count, ggml_backend_dev_name(devs[i]));
         printf("  Device description: %s\n", ggml_backend_dev_description(devs[i]));

tools/imatrix/imatrix.cpp

@@ -24,7 +24,8 @@ static void print_usage(int, char ** argv) {
     LOG("\n    %s \\\n"
             "       -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] \\\n"
             "       [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n"
-            "       [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]\n" , argv[0]);
+            "       [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...] \\\n"
+            "       [--parse-special]\n" , argv[0]);
     LOG("\n");
 }
 
@@ -439,7 +440,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
     auto tim1 = std::chrono::high_resolution_clock::now();
     LOG_INF("%s: tokenizing the input ..\n", __func__);
 
-    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);
+    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true, params.parse_special);
 
     auto tim2 = std::chrono::high_resolution_clock::now();
     LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());

tools/llama-bench/llama-bench.cpp

@@ -219,6 +219,7 @@ struct cmd_params {
     std::vector<std::vector<llama_model_tensor_buft_override>> tensor_buft_overrides;
     std::vector<bool>                use_mmap;
     std::vector<bool>                embeddings;
+    std::vector<bool>                no_op_offload;
     ggml_numa_strategy               numa;
     int                              reps;
     ggml_sched_priority              prio;
@@ -253,6 +254,7 @@ static const cmd_params cmd_params_defaults = {
     /* tensor_buft_overrides*/ { std::vector<llama_model_tensor_buft_override>{{nullptr,nullptr}} },
     /* use_mmap             */ { true },
     /* embeddings           */ { false },
+    /* no_op_offload        */ { false },
     /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
     /* reps                 */ 5,
     /* prio                 */ GGML_SCHED_PRIO_NORMAL,
@@ -311,6 +313,7 @@ static void print_usage(int /* argc */, char ** argv) {
            join(cmd_params_defaults.embeddings, ",").c_str());
     printf("  -ts, --tensor-split <ts0/ts1/..>          (default: 0)\n");
     printf("  -ot --override-tensors <tensor name pattern>=<buffer type>;... (default: disabled)\n");
+    printf("  -nopo, --no-op-offload <i>                (default: 0)\n");
     printf("  -r, --repetitions <n>                     (default: %d)\n", cmd_params_defaults.reps);
     printf("  --prio <0|1|2|3>                          (default: %d)\n", cmd_params_defaults.prio);
     printf("  --delay <0...N> (seconds)                 (default: %d)\n", cmd_params_defaults.delay);
@@ -588,6 +591,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = string_split<bool>(argv[i], split_delim);
             params.embeddings.insert(params.embeddings.end(), p.begin(), p.end());
+        } else if (arg == "-nopo" || arg == "--no-op-offload") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<bool>(argv[i], split_delim);
+            params.no_op_offload.insert(params.no_op_offload.end(), p.begin(), p.end());
         } else if (arg == "-ts" || arg == "--tensor-split") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -794,6 +804,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.embeddings.empty()) {
         params.embeddings = cmd_params_defaults.embeddings;
     }
+    if (params.no_op_offload.empty()) {
+        params.no_op_offload = cmd_params_defaults.no_op_offload;
+    }
     if (params.n_threads.empty()) {
         params.n_threads = cmd_params_defaults.n_threads;
     }
@@ -833,6 +846,7 @@ struct cmd_params_instance {
     std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
     bool               use_mmap;
     bool               embeddings;
+    bool               no_op_offload;
 
     llama_model_params to_llama_mparams() const {
         llama_model_params mparams = llama_model_default_params();
@@ -902,6 +916,7 @@ struct cmd_params_instance {
         cparams.offload_kqv = !no_kv_offload;
         cparams.flash_attn  = flash_attn;
         cparams.embeddings  = embeddings;
+        cparams.op_offload  = !no_op_offload;
 
         return cparams;
     }
@@ -921,6 +936,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & ot : params.tensor_buft_overrides)
     for (const auto & mmp : params.use_mmap)
     for (const auto & embd : params.embeddings)
+    for (const auto & nopo : params.no_op_offload)
     for (const auto & nb : params.n_batch)
     for (const auto & nub : params.n_ubatch)
     for (const auto & tk : params.type_k)
@@ -959,6 +975,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
+                /* .no_op_offload= */ nopo,
             };
             instances.push_back(instance);
         }
@@ -990,6 +1007,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
+                /* .no_op_offload= */ nopo,
             };
             instances.push_back(instance);
         }
@@ -1021,6 +1039,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
+                /* .no_op_offload= */ nopo,
             };
             instances.push_back(instance);
         }
@@ -1056,6 +1075,7 @@ struct test {
     std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
     bool                     use_mmap;
     bool                     embeddings;
+    bool                     no_op_offload;
     int                      n_prompt;
     int                      n_gen;
     int                      n_depth;
@@ -1089,6 +1109,7 @@ struct test {
         tensor_buft_overrides = inst.tensor_buft_overrides;
         use_mmap       = inst.use_mmap;
         embeddings     = inst.embeddings;
+        no_op_offload  = inst.no_op_offload;
         n_prompt       = inst.n_prompt;
         n_gen          = inst.n_gen;
         n_depth        = inst.n_depth;
@@ -1134,7 +1155,7 @@ struct test {
             "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
             "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
             "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
-            "use_mmap",     "embeddings",   "n_prompt",       "n_gen",      "n_depth",      "test_time",
+            "use_mmap",     "embeddings",   "no_op_offload",   "n_prompt",       "n_gen",      "n_depth",      "test_time",
             "avg_ns",       "stddev_ns",    "avg_ts",         "stddev_ts",
         };
         return fields;
@@ -1146,7 +1167,7 @@ struct test {
         if (field == "build_number" || field == "n_batch" || field == "n_ubatch" || field == "n_threads" ||
             field == "poll" || field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" ||
             field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "n_depth" ||
-            field == "avg_ns" || field == "stddev_ns") {
+            field == "avg_ns" || field == "stddev_ns" || field == "no_op_offload") {
             return INT;
         }
         if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" || field == "flash_attn" ||
@@ -1222,6 +1243,7 @@ struct test {
                                             tensor_buft_overrides_str,
                                             std::to_string(use_mmap),
                                             std::to_string(embeddings),
+                                            std::to_string(no_op_offload),
                                             std::to_string(n_prompt),
                                             std::to_string(n_gen),
                                             std::to_string(n_depth),
@@ -1404,6 +1426,9 @@ struct markdown_printer : public printer {
         if (field == "test") {
             return 15;
         }
+        if (field == "no_op_offload") {
+            return 4;
+        }
 
         int width = std::max((int) field.length(), 10);
 
@@ -1435,6 +1460,9 @@ struct markdown_printer : public printer {
         if (field == "embeddings") {
             return "embd";
         }
+        if (field == "no_op_offload") {
+            return "nopo";
+        }
         if (field == "tensor_split") {
             return "ts";
         }
@@ -1503,6 +1531,9 @@ struct markdown_printer : public printer {
         if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) {
             fields.emplace_back("embeddings");
         }
+        if (params.no_op_offload.size() > 1 || params.no_op_offload != cmd_params_defaults.no_op_offload) {
+            fields.emplace_back("no_op_offload");
+        }
         fields.emplace_back("test");
         fields.emplace_back("t/s");
 

tools/main/main.cpp

@@ -152,7 +152,12 @@ int main(int argc, char ** argv) {
 
     LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
 
-    auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
+    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (!cpu_dev) {
+        LOG_ERR("%s: no CPU backend found\n", __func__);
+        return 1;
+    }
+    auto * reg = ggml_backend_dev_backend_reg(cpu_dev);
     auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_new");
     auto * ggml_threadpool_free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_free");
 

tools/mtmd/CMakeLists.txt

@@ -28,6 +28,7 @@ endif()
 
 add_library(mtmd OBJECT
             mtmd.cpp
+            mtmd-helper.cpp
             mtmd.h
             clip.cpp
             clip.h

tools/mtmd/README.md

@@ -16,38 +16,7 @@ The naming and structure related to multimodal support have evolved, which might
 
 ## Pre-quantized models
 
-These are ready-to-use models, most of them come with `Q4_K_M` quantization by default:
-
-```sh
-# Gemma 3
-llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
-llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
-llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
-
-# SmolVLM
-llama-mtmd-cli -hf ggml-org/SmolVLM-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM-256M-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM-500M-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
-
-# Pixtral 12B
-llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF
-
-# Qwen 2 VL
-llama-mtmd-cli -hf ggml-org/Qwen2-VL-2B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/Qwen2-VL-7B-Instruct-GGUF
-
-# Qwen 2.5 VL
-llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-3B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-7B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-32B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-72B-Instruct-GGUF
-
-# Mistral Small 3.1 24B (IQ2_M quantization)
-llama-mtmd-cli -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF --chat-template mistral-v7
-```
+See the list of pre-quantized model [here](../../docs/multimodal.md)
 
 ## How it works and what is `mmproj`?
 
@@ -79,6 +48,7 @@ For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` fla
 - [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint
 - Qwen 2 VL and Qwen 2.5 VL (from [Qwen](https://huggingface.co/Qwen))
 - [Mistral Small 3.1 24B](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503)
+- InternVL 2.5 and InternVL 3 from [OpenGVLab](https://huggingface.co/OpenGVLab) (note: we don't support conversion of `InternVL3-*-hf` model, only non-HF version is supported ; `InternLM2Model` **text** model is not supported)
 
 For older models, please refer to the relevant guide for instructions on how to obtain or create them:
 

tools/mtmd/clip-impl.h

@@ -33,9 +33,6 @@
 #define KEY_PROJ_TYPE           "clip.projector_type"
 #define KEY_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"
 
-#define KEY_USE_GLU_MLP         "clip.use_glu_mlp"  // for qwen2.5vl
-#define KEY_USE_RMS_NORM        "clip.use_rms_norm" // for qwen2.5vl
-
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
 #define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
@@ -56,12 +53,16 @@
 #define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
 #define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
 #define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
+#define TN_ATTN_K_NORM     "%s.blk.%d.attn_k_norm.%s"
+#define TN_ATTN_Q_NORM     "%s.blk.%d.attn_q_norm.%s"
 #define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
 #define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
 #define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
-#define TN_LN_1            "%s.blk.%d.ln1.%s"
-#define TN_LN_2            "%s.blk.%d.ln2.%s"
+#define TN_LN_1            "%s.blk.%d.ln1.%s" // layer norm
+#define TN_LN_2            "%s.blk.%d.ln2.%s" // layer norm
+#define TN_LS_1            "%s.blk.%d.ls1.%s" // layer scale
+#define TN_LS_2            "%s.blk.%d.ls2.%s" // layer scale
 #define TN_LN_PRE          "%s.pre_ln.%s"
 #define TN_LN_POST         "%s.post_ln.%s"
 #define TN_LLAVA_PROJ      "mm.%d.%s"
@@ -93,6 +94,9 @@
 #define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
 #define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
 
+// align x to upper multiple of n
+#define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
+
 enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
@@ -105,6 +109,7 @@ enum projector_type {
     PROJECTOR_TYPE_IDEFICS3,
     PROJECTOR_TYPE_PIXTRAL,
     PROJECTOR_TYPE_QWEN25VL,
+    PROJECTOR_TYPE_INTERNVL,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -119,6 +124,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
     { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
     { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
+    { PROJECTOR_TYPE_INTERNVL,  "internvl"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

tools/mtmd/clip.cpp

@@ -174,6 +174,10 @@ struct clip_hparams {
     int32_t n_layer;
     int32_t proj_scale_factor = 0; // idefics3
 
+    // for models using dynamic image size, we need to have a smaller image size to warmup
+    // otherwise, user will get OOM everytime they load the model
+    int32_t warmup_image_size = 0;
+
     ffn_op_type ffn_op = FFN_GELU;
 
     patch_merge_type mm_patch_merge_type = PATCH_MERGE_FLAT;
@@ -201,6 +205,9 @@ struct clip_layer {
     ggml_tensor * o_w = nullptr;
     ggml_tensor * o_b = nullptr;
 
+    ggml_tensor * k_norm = nullptr;
+    ggml_tensor * q_norm = nullptr;
+
     // layernorm 1
     ggml_tensor * ln_1_w = nullptr;
     ggml_tensor * ln_1_b = nullptr;
@@ -215,6 +222,10 @@ struct clip_layer {
     // layernorm 2
     ggml_tensor * ln_2_w = nullptr;
     ggml_tensor * ln_2_b = nullptr;
+
+    // layer scale (no bias)
+    ggml_tensor * ls_1_w = nullptr;
+    ggml_tensor * ls_2_w = nullptr;
 };
 
 struct clip_vision_model {
@@ -352,9 +363,12 @@ struct clip_ctx {
 
     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_cpu) {
+            throw std::runtime_error("failed to initialize CPU backend");
+        }
+        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));
@@ -369,7 +383,7 @@ struct clip_ctx {
         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)
+            ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false, true)
         );
     }
 
@@ -586,6 +600,9 @@ struct clip_graph {
 
     // Qwen2VL and Qwen2.5VL use M-RoPE
     ggml_cgraph * build_qwen2vl() {
+        GGML_ASSERT(model.patch_bias == nullptr);
+        GGML_ASSERT(model.class_embedding == nullptr);
+
         const int batch_size       = 1;
         const bool use_window_attn = hparams.n_wa_pattern > 0;
         const int n_wa_pattern     = hparams.n_wa_pattern;
@@ -622,10 +639,6 @@ struct clip_graph {
                 n_embd, n_patches_x * n_patches_y, batch_size);
         }
 
-        if (model.patch_bias) {
-            inp = ggml_add(ctx0, inp, model.patch_bias);
-        }
-
         ggml_tensor * inpL           = inp;
         ggml_tensor * window_mask    = nullptr;
         ggml_tensor * window_idx     = nullptr;
@@ -856,6 +869,67 @@ struct clip_graph {
         return gf;
     }
 
+    ggml_cgraph * build_internvl() {
+        GGML_ASSERT(model.class_embedding != nullptr);
+        GGML_ASSERT(model.position_embeddings != nullptr);
+
+        const int n_pos = n_patches + 1;
+        ggml_tensor * inp = build_inp();
+
+        // add CLS token
+        inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+
+        ggml_tensor * cur = build_vit(
+                                inp, n_pos,
+                                NORM_TYPE_NORMAL,
+                                hparams.ffn_op,
+                                model.position_embeddings,
+                                nullptr);
+
+        // remove CLS token
+        cur = ggml_view_2d(ctx0, cur,
+            n_embd, n_patches,
+            ggml_row_size(cur->type, n_embd), 0);
+
+        // pixel shuffle
+        {
+            const int scale_factor = model.hparams.proj_scale_factor;
+            const int bsz    = 1; // batch size, always 1 for now since we don't support batching
+            const int height = n_patches_y;
+            const int width  = n_patches_x;
+            GGML_ASSERT(scale_factor > 0);
+            cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, height / scale_factor, width, bsz);
+            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
+                n_embd * scale_factor * scale_factor,
+                height / scale_factor,
+                width / scale_factor,
+                bsz);
+            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+            // flatten to 2D
+            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur),
+                n_embd * scale_factor * scale_factor,
+                cur->ne[1] * cur->ne[2]);
+        }
+
+        // projector (always using GELU activation)
+        {
+            // projector LayerNorm uses pytorch's default eps = 1e-5
+            // ref: https://huggingface.co/OpenGVLab/InternVL3-8B-Instruct/blob/a34d3e4e129a5856abfd6aa6de79776484caa14e/modeling_internvl_chat.py#L79
+            cur = build_norm(cur, model.mm_0_w, model.mm_0_b, NORM_TYPE_NORMAL, 1e-5, -1);
+            cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
+            cur = ggml_add(ctx0, cur, model.mm_1_b);
+            cur = ggml_gelu(ctx0, cur);
+            cur = ggml_mul_mat(ctx0, model.mm_3_w, cur);
+            cur = ggml_add(ctx0, cur, model.mm_3_b);
+        }
+
+        // build the graph
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     // this graph is used by llava, granite and glm
     // due to having embedding_stack (used by granite), we cannot reuse build_vit
     ggml_cgraph * build_llava() {
@@ -887,10 +961,6 @@ struct clip_graph {
 
         ggml_tensor * inp = build_inp();
 
-        if (model.patch_bias) {
-            inp = ggml_add(ctx0, inp, model.patch_bias);
-        }
-
         // concat class_embeddings and patch_embeddings
         if (model.class_embedding) {
             inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
@@ -1257,11 +1327,6 @@ private:
                 ggml_tensor * learned_pos_embd,
                 std::function<ggml_tensor *(ggml_tensor *, const clip_layer &)> add_pos
             ) {
-        if (model.patch_bias) {
-            inp = ggml_add(ctx0, inp, model.patch_bias);
-            cb(inp, "patch_bias", -1);
-        }
-
         if (learned_pos_embd) {
             inp = ggml_add(ctx0, inp, learned_pos_embd);
             cb(inp, "pos_embed", -1);
@@ -1301,6 +1366,16 @@ private:
                     Vcur = ggml_add(ctx0, Vcur, layer.v_b);
                 }
 
+                if (layer.q_norm) {
+                    Qcur = build_norm(Qcur, layer.q_norm, NULL, norm_t, eps, il);
+                    cb(Qcur, "Qcur_norm", il);
+                }
+
+                if (layer.k_norm) {
+                    Kcur = build_norm(Kcur, layer.k_norm, NULL, norm_t, eps, il);
+                    cb(Kcur, "Kcur_norm", il);
+                }
+
                 Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
                 Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos);
                 Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos);
@@ -1321,6 +1396,11 @@ private:
                 cb(cur, "attn_out", il);
             }
 
+            if (layer.ls_1_w) {
+                cur = ggml_mul(ctx0, cur, layer.ls_1_w);
+                cb(cur, "attn_out_scaled", il);
+            }
+
             // re-add the layer input, e.g., residual
             cur = ggml_add(ctx0, cur, inpL);
 
@@ -1341,6 +1421,11 @@ private:
 
             cb(cur, "ffn_out", il);
 
+            if (layer.ls_2_w) {
+                cur = ggml_mul(ctx0, cur, layer.ls_2_w);
+                cb(cur, "ffn_out_scaled", il);
+            }
+
             // residual 2
             cur = ggml_add(ctx0, inpL, cur);
             cb(cur, "layer_out", il);
@@ -1362,6 +1447,10 @@ private:
         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, n_patches, n_embd);
         inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
+        if (model.patch_bias) {
+            inp = ggml_add(ctx0, inp, model.patch_bias);
+            cb(inp, "patch_bias", -1);
+        }
         return inp;
     }
 
@@ -1624,6 +1713,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             {
                 res = graph.build_minicpmv();
             } break;
+        case PROJECTOR_TYPE_INTERNVL:
+            {
+                res = graph.build_internvl();
+            } break;
         default:
             {
                 res = graph.build_llava();
@@ -1720,6 +1813,9 @@ struct clip_model_loader {
             get_u32(KEY_IMAGE_CROP_RESOLUTION,    hparams.image_crop_resolution, false);
             get_arr_int(KEY_IMAGE_GRID_PINPOINTS, hparams.image_grid_pinpoints, false);
 
+            // default warmup value
+            hparams.warmup_image_size = hparams.image_size;
+
             ctx_clip.has_llava_projector = ctx_clip.proj_type == PROJECTOR_TYPE_MLP
                                         || ctx_clip.proj_type == PROJECTOR_TYPE_MLP_NORM
                                         || ctx_clip.proj_type == PROJECTOR_TYPE_LDP
@@ -1787,12 +1883,14 @@ struct clip_model_loader {
                         }
                     } break;
                 case PROJECTOR_TYPE_IDEFICS3:
+                case PROJECTOR_TYPE_INTERNVL:
                     {
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
                     } break;
                 case PROJECTOR_TYPE_PIXTRAL:
                     {
                         hparams.rope_theta = 10000.0f;
+                        hparams.warmup_image_size = hparams.patch_size * 8;
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false);
                     } break;
                 case PROJECTOR_TYPE_GEMMA3:
@@ -1803,8 +1901,19 @@ struct clip_model_loader {
                         // test model (tinygemma3) has a different value, we optionally read it
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
                     } break;
+                case PROJECTOR_TYPE_QWEN2VL:
+                    {
+                        // max image size = sqrt(max_pixels)
+                        // https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/blob/main/preprocessor_config.json
+                        hparams.image_size = 3584;
+                        hparams.warmup_image_size = hparams.patch_size * 8;
+                    } break;
                 case PROJECTOR_TYPE_QWEN25VL:
                     {
+                        // max image size = sqrt(max_pixels)
+                        // https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/blob/main/preprocessor_config.json
+                        hparams.image_size = 3584;
+                        hparams.warmup_image_size = hparams.patch_size * 8;
                         get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern);
                     } break;
                 default:
@@ -1892,8 +2001,13 @@ struct clip_model_loader {
             layer.q_w    = get_tensor(string_format(TN_ATTN_Q,      "v", il, "weight"));
             layer.v_w    = get_tensor(string_format(TN_ATTN_V,      "v", il, "weight"));
             layer.o_w    = get_tensor(string_format(TN_ATTN_OUTPUT, "v", il, "weight"));
+            layer.k_norm = get_tensor(string_format(TN_ATTN_K_NORM, "v", il, "weight"), false);
+            layer.q_norm = get_tensor(string_format(TN_ATTN_Q_NORM, "v", il, "weight"), false);
             layer.ln_1_w = get_tensor(string_format(TN_LN_1,        "v", il, "weight"), false);
             layer.ln_2_w = get_tensor(string_format(TN_LN_2,        "v", il, "weight"), false);
+            layer.ls_1_w = get_tensor(string_format(TN_LS_1,        "v", il, "weight"), false); // no bias
+            layer.ls_2_w = get_tensor(string_format(TN_LS_2,        "v", il, "weight"), false); // no bias
+
             layer.k_b    = get_tensor(string_format(TN_ATTN_K,      "v", il, "bias"), false);
             layer.q_b    = get_tensor(string_format(TN_ATTN_Q,      "v", il, "bias"), false);
             layer.v_b    = get_tensor(string_format(TN_ATTN_V,      "v", il, "bias"), false);
@@ -1901,7 +2015,7 @@ struct clip_model_loader {
             layer.ln_1_b = get_tensor(string_format(TN_LN_1,        "v", il, "bias"), false);
             layer.ln_2_b = get_tensor(string_format(TN_LN_2,        "v", il, "bias"), false);
 
-            // new naming
+            // ffn
             layer.ff_up_w   = get_tensor(string_format(TN_FFN_UP,   "v", il, "weight"));
             layer.ff_up_b   = get_tensor(string_format(TN_FFN_UP,   "v", il, "bias"),   false);
             layer.ff_gate_w = get_tensor(string_format(TN_FFN_GATE, "v", il, "weight"), false);
@@ -2049,6 +2163,15 @@ struct clip_model_loader {
                     vision_model.mm_input_norm_w   = get_tensor(TN_MM_INP_NORM,     false);
                     vision_model.mm_patch_merger_w = get_tensor(TN_MM_PATCH_MERGER, false);
                 } break;
+            case PROJECTOR_TYPE_INTERNVL:
+                {
+                    vision_model.mm_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
+                    vision_model.mm_0_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "bias"));
+                    vision_model.mm_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
+                    vision_model.mm_1_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "bias"));
+                    vision_model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
+                    vision_model.mm_3_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "bias"));
+                } break;
             default:
                 GGML_ASSERT(false && "unknown projector type");
         }
@@ -2096,13 +2219,14 @@ struct clip_model_loader {
         // create a fake batch
         clip_image_f32_batch batch;
         clip_image_f32_ptr img(clip_image_f32_init());
-        img->nx = ctx_clip.vision_model.hparams.image_size;
-        img->ny = ctx_clip.vision_model.hparams.image_size;
+        img->nx = ctx_clip.vision_model.hparams.warmup_image_size;
+        img->ny = ctx_clip.vision_model.hparams.warmup_image_size;
         img->buf.resize(img->nx * img->ny * 3);
         batch.entries.push_back(std::move(img));
 
         ggml_cgraph * gf = clip_image_build_graph(&ctx_clip, batch);
         ggml_backend_sched_reserve(ctx_clip.sched.get(), gf);
+
         for (size_t i = 0; i < ctx_clip.backend_ptrs.size(); ++i) {
             ggml_backend_t backend = ctx_clip.backend_ptrs[i];
             ggml_backend_buffer_type_t buft = ctx_clip.backend_buft[i];
@@ -2185,9 +2309,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity) {
 
 struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params) {
     g_logger_state.verbosity_thold = ctx_params.verbosity;
-    clip_ctx * ctx_clip = new clip_ctx(ctx_params);
+    clip_ctx * ctx_clip = nullptr;
 
     try {
+        ctx_clip = new clip_ctx(ctx_params);
         clip_model_loader loader(fname, *ctx_clip);
         loader.load_hparams();
         loader.load_tensors();
@@ -2500,8 +2625,8 @@ struct image_manipulation {
         float target_width_f  = static_cast<float>(inp_size.width)  * scale;
         float target_height_f = static_cast<float>(inp_size.height) * scale;
 
-        int aligned_width  = GGML_PAD((int)target_width_f,  align_size);
-        int aligned_height = GGML_PAD((int)target_height_f, align_size);
+        int aligned_width  = CLIP_ALIGN((int)target_width_f,  align_size);
+        int aligned_height = CLIP_ALIGN((int)target_height_f, align_size);
 
         return {aligned_width, aligned_height};
     }
@@ -2820,10 +2945,9 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
     }
     else if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) {
         clip_image_u8 resized;
-        auto patch_size = clip_get_patch_size(ctx) * 2;
-        int nx = ceil((float)img->nx / patch_size) * patch_size;
-        int ny = ceil((float)img->ny / patch_size) * patch_size;
-        image_manipulation::bicubic_resize(*img, resized, nx, ny);
+        auto patch_size = params.patch_size * 2;
+        auto new_size = image_manipulation::calc_size_preserved_ratio(original_size, patch_size, params.image_size);
+        image_manipulation::bicubic_resize(*img, resized, new_size.width, new_size.height);
 
         clip_image_f32_ptr img_f32(clip_image_f32_init());
         // clip_image_f32_ptr res(clip_image_f32_init());
@@ -2834,7 +2958,9 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
     }
     else if (ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE
             || ctx->proj_type == PROJECTOR_TYPE_GEMMA3
-            || ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) {
+            || ctx->proj_type == PROJECTOR_TYPE_IDEFICS3
+            || ctx->proj_type == PROJECTOR_TYPE_INTERNVL // TODO @ngxson : support dynamic resolution
+    ) {
         clip_image_u8 resized_image;
         int sz = params.image_size;
         image_manipulation::resize_and_pad_image(*img, resized_image, {sz, sz});
@@ -2984,9 +3110,13 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
 
     int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
 
-    if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2 || ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP
+            || ctx->proj_type == PROJECTOR_TYPE_LDPV2
+            || ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE) {
         n_patches /= 4;
-        n_patches += 2; // for BOI and EOI token embeddings
+        if (ctx->vision_model.mm_glm_tok_boi) {
+            n_patches += 2; // for BOI and EOI token embeddings
+        }
     } else if (ctx->proj_type == PROJECTOR_TYPE_MINICPMV) {
         if (ctx->minicpmv_version == 2) {
             n_patches = 96;
@@ -3009,7 +3139,8 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         int n_per_side = params.image_size / params.patch_size;
         int n_per_side_2d_pool = n_per_side / params.proj_scale_factor;
         n_patches = n_per_side_2d_pool * n_per_side_2d_pool;
-    } else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) {
+    } else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3 || ctx->proj_type == PROJECTOR_TYPE_INTERNVL) {
+        // both W and H are divided by proj_scale_factor
         n_patches /= (params.proj_scale_factor * params.proj_scale_factor);
     } else if (ctx->proj_type == PROJECTOR_TYPE_PIXTRAL) {
         int n_merge = params.spatial_merge_size;
@@ -3404,6 +3535,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             } break;
         case PROJECTOR_TYPE_GEMMA3:
         case PROJECTOR_TYPE_IDEFICS3:
+        case PROJECTOR_TYPE_INTERNVL:
             {
                 // do nothing
             } break;
@@ -3430,6 +3562,14 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     // the last node is the embedding tensor
     ggml_tensor * embeddings = ggml_graph_node(gf, -1);
 
+    // sanity check (only support batch size of 1 for now)
+    const int n_tokens_out = embeddings->ne[1];
+    const int expected_n_tokens_out = clip_n_output_tokens(ctx, imgs.entries[0].get());
+    if (n_tokens_out != expected_n_tokens_out) {
+        LOG_ERR("%s: expected %d tokens, got %d\n", __func__, expected_n_tokens_out, n_tokens_out);
+        GGML_ABORT("Invalid number of output tokens");
+    }
+
     // copy the embeddings to the location passed by the user
     ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
 
@@ -3600,6 +3740,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->vision_model.mm_input_proj_w->ne[0];
         case PROJECTOR_TYPE_IDEFICS3:
             return ctx->vision_model.projection->ne[1];
+        case PROJECTOR_TYPE_INTERNVL:
+            return ctx->vision_model.mm_3_w->ne[1];
         default:
             GGML_ABORT("Unknown projector type");
     }

tools/mtmd/llava.cpp

@@ -212,6 +212,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     ggml_build_forward_expand(gf, flatten);
 
     ggml_backend_ptr backend { ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr) };
+    GGML_ASSERT(backend != nullptr && "failed to initialize CPU backend");
     ggml_backend_graph_compute(backend.get(), gf);
 
     struct ggml_tensor* result = ggml_graph_node(gf, -1);

tools/mtmd/mtmd-helper.cpp

@@ -0,0 +1,310 @@
+#include "mtmd.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cinttypes>
+#include <vector>
+
+#define LOG_INF(...) fprintf(stdout, __VA_ARGS__)
+#define LOG_ERR(...) fprintf(stderr, __VA_ARGS__)
+
+size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks) {
+    size_t n_tokens = 0;
+    for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
+        auto chunk = mtmd_input_chunks_get(chunks, i);
+        auto chunk_type = mtmd_input_chunk_get_type(chunk);
+        if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            size_t n_tokens_text;
+            mtmd_input_chunk_get_tokens_text(chunk, &n_tokens_text);
+            n_tokens += n_tokens_text;
+        } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            auto tokens_image = mtmd_input_chunk_get_tokens_image(chunk);
+            n_tokens += mtmd_image_tokens_get_n_tokens(tokens_image);
+        } else {
+            GGML_ASSERT(false && "chunk type not supported");
+        }
+    }
+    return n_tokens;
+}
+
+llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks) {
+    llama_pos n_pos = 0;
+    for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
+        auto chunk = mtmd_input_chunks_get(chunks, i);
+        auto chunk_type = mtmd_input_chunk_get_type(chunk);
+        if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            size_t n_tokens_text;
+            mtmd_input_chunk_get_tokens_text(chunk, &n_tokens_text);
+            n_pos += n_tokens_text;
+        } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            auto tokens_image = mtmd_input_chunk_get_tokens_image(chunk);
+            n_pos += mtmd_image_tokens_get_n_pos(tokens_image);
+        } else {
+            GGML_ASSERT(false && "chunk type not supported");
+        }
+    }
+    return n_pos;
+}
+
+// helper struct to make working with embd batch easier
+// note: this will be removed after llama_batch_ext refactoring
+struct decode_embd_batch {
+    int n_pos_per_embd;
+    int n_mmproj_embd;
+    std::vector<llama_pos>      pos;
+    std::vector<llama_pos>      pos_view; // used by mrope
+    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, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+        pos     .resize(n_tokens * n_pos_per_embd);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        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(),
+        };
+    }
+
+    void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {
+        seq_id_0[0] = seq_id;
+        for (int i = 0; i < batch.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;
+        }
+    }
+
+    void set_position_mrope(llama_pos pos_0, int nx, int ny, llama_seq_id seq_id) {
+        GGML_ASSERT(n_pos_per_embd == 4);
+        seq_id_0[0] = seq_id;
+        for (int y = 0; y < ny; y++) {
+            for (int x = 0; x < nx; x++) {
+                int i = y * nx + x;
+                pos[i                     ] = pos_0;
+                pos[i + batch.n_tokens    ] = pos_0 + y;
+                pos[i + batch.n_tokens * 2] = pos_0 + x;
+                pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
+            }
+        }
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    llama_batch get_view(int offset, int n_tokens) {
+        llama_pos * pos_ptr;
+        pos_view.clear();
+        pos_view.reserve(n_tokens * n_pos_per_embd);
+        if (n_pos_per_embd > 1) {
+            // mrope
+            // for example, with layout of src: 1234...1234...1234...1234...
+            //       offset 2 will give us dst: 34...34...34...34...
+            for (int i = 0; i < n_pos_per_embd; i++) {
+                // assume n_tokens is less than or equal to batch.n_tokens
+                // batch.n_tokens is number of **total** tokens
+                // n_tokens is number of viewed token
+                size_t src_idx = i * batch.n_tokens + offset;
+                pos_view.insert(pos_view.end(),
+                    pos.data() + src_idx,
+                    pos.data() + src_idx + n_tokens);
+            }
+            pos_ptr = pos_view.data();
+        } else {
+            // normal
+            pos_ptr = pos.data() + offset;
+        }
+        return {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ batch.embd     + offset * n_mmproj_embd,
+            /*pos            =*/ pos_ptr,
+            /*n_seq_id       =*/ batch.n_seq_id + offset,
+            /*seq_id         =*/ batch.seq_id   + offset,
+            /*logits         =*/ batch.logits   + offset,
+        };
+    }
+};
+
+// Helper function for decoding an image whose embeddings have already been calculated
+int32_t mtmd_helper_decode_image_chunk(
+        mtmd_context * ctx,
+        struct llama_context * lctx,
+        const mtmd_input_chunk * chunk,
+        float * encoded_embd,
+        llama_pos n_past,
+        llama_seq_id seq_id,
+        int32_t n_batch,
+        llama_pos * new_n_past) {
+    if (mtmd_input_chunk_get_type(chunk) != MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        LOG_ERR("failed to decode image chunk: input chunk not of image type\n");
+        return -1;
+    }
+    const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk);
+    if (!image_tokens) {
+        LOG_ERR("failed to decode image chunk: image tokens are null\n");
+        return -1;
+    }
+
+    const llama_model * model = llama_get_model(lctx);
+    int n_mmproj_embd = llama_model_n_embd(model);
+    int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;
+
+    int32_t n_tokens = mtmd_image_tokens_get_n_tokens(image_tokens);
+    int32_t i_batch = 0;
+    int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
+    decode_embd_batch batch_embd(encoded_embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
+
+    const int nx = mtmd_image_tokens_get_nx(image_tokens);
+    const int ny = mtmd_image_tokens_get_ny(image_tokens);
+
+    if (mtmd_decode_use_mrope(ctx)) {
+        batch_embd.set_position_mrope(n_past, nx, ny, seq_id);
+    } else {
+        batch_embd.set_position_normal(n_past, seq_id);
+    }
+
+    if (mtmd_decode_use_non_causal(ctx)) {
+        llama_set_causal_attn(lctx, false);
+        // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
+    }
+
+    while (i_batch < n_img_batches) { // split into batches
+        int pos_offset = i_batch*n_batch;
+        int n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
+        llama_batch batch_embd_view = batch_embd.get_view(pos_offset, n_tokens_batch);
+
+        LOG_INF("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
+
+        int64_t t1 = ggml_time_ms();
+        int32_t ret = llama_decode(lctx, batch_embd_view);
+        if (ret != 0) {
+            LOG_ERR("failed to decode image\n");
+            llama_set_causal_attn(lctx, true); // restore causal attn
+            return ret;
+        }
+
+        LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1);
+
+        i_batch++;
+    }
+
+    n_past += mtmd_image_tokens_get_n_pos(image_tokens);
+    *new_n_past = n_past;
+
+    if (mtmd_decode_use_non_causal(ctx)) {
+        llama_set_causal_attn(lctx, true);
+    }
+    return 0;
+}
+
+int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
+        struct llama_context * lctx,
+        const mtmd_input_chunk * chunk,
+        llama_pos n_past,
+        llama_seq_id seq_id,
+        int32_t n_batch,
+        bool logits_last,
+        llama_pos * new_n_past) {
+    int32_t ret;
+    llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
+    auto chunk_type = mtmd_input_chunk_get_type(chunk);
+
+    if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        size_t n_tokens;
+        const auto tokens = mtmd_input_chunk_get_tokens_text(chunk, &n_tokens);
+        // LOG_INF("decoding text chunk, n_tokens = %zu\n", n_tokens);
+        size_t i = 0;
+        while (i < n_tokens) { // split into batches
+            text_batch.n_tokens = 0; // clear the batch
+            for (; i < n_tokens && text_batch.n_tokens < n_batch; i++) {
+                text_batch.n_tokens++;
+                text_batch.token   [i]    = tokens[i];
+                text_batch.pos     [i]    = n_past++;
+                text_batch.n_seq_id[i]    = 1;
+                text_batch.seq_id  [i][0] = seq_id;
+                text_batch.logits  [i]    = false;
+            }
+            bool is_last_token = (i == n_tokens);
+            if (logits_last && is_last_token) {
+                text_batch.logits[text_batch.n_tokens - 1] = true;
+            }
+            ret = llama_decode(lctx, text_batch);
+            if (ret != 0) {
+                LOG_ERR("failed to decode text\n");
+                llama_batch_free(text_batch);
+                return ret;
+            }
+            *new_n_past += text_batch.n_tokens;
+        }
+
+    } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk);
+        int64_t t0 = ggml_time_ms();
+
+        LOG_INF("encoding image or slice...\n");
+
+        ret = mtmd_encode(ctx, image_tokens);
+        if (ret != 0) {
+            LOG_ERR("failed to encode image\n");
+            llama_batch_free(text_batch);
+            return ret;
+        }
+
+        LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
+
+        float * embd = mtmd_get_output_embd(ctx);
+        ret = mtmd_helper_decode_image_chunk(ctx, lctx, chunk, embd, n_past, seq_id, n_batch, new_n_past);
+        if (ret != 0) {
+            LOG_ERR("failed to decode image\n");
+            llama_batch_free(text_batch);
+            return ret;
+        }
+    } else {
+        GGML_ABORT("chunk type not supported");
+    }
+
+    return 0;
+}
+
+int32_t mtmd_helper_eval_chunks(mtmd_context * ctx,
+                                struct llama_context * lctx,
+                                const mtmd_input_chunks * chunks,
+                                llama_pos n_past,
+                                llama_seq_id seq_id,
+                                int32_t n_batch,
+                                bool logits_last,
+                                llama_pos * new_n_past) {
+    size_t n_chunks = mtmd_input_chunks_size(chunks);
+    if (n_chunks == 0) {
+        LOG_ERR("no chunks to eval\n");
+        return 0;
+    }
+
+    for (size_t i = 0; i < n_chunks; i++) {
+        bool chunk_logits_last = (i == n_chunks - 1) && logits_last;
+        auto chunk = mtmd_input_chunks_get(chunks, i);
+
+        int32_t res = mtmd_helper_eval_chunk_single(ctx, lctx, chunk, n_past, seq_id, n_batch, chunk_logits_last, &n_past);
+        if (res != 0) {
+            LOG_ERR("failed to eval chunk %zu\n", i);
+            return res;
+        }
+        *new_n_past = n_past;
+    }
+
+    return 0;
+}

tools/mtmd/mtmd.cpp

@@ -252,6 +252,13 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
 
     }
 
+    else if (proj_type == PROJECTOR_TYPE_INTERNVL) {
+        // <img> ... (image embeddings) ... </img>
+        marker_modified = "<img>" + ctx->image_marker + "</img>";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+
+    }
+
     // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
     // for glm-edge, BOI and EOI token's embeddings are not present in the text model
 
@@ -454,275 +461,26 @@ float * mtmd_get_output_embd(mtmd_context * ctx) {
     return ctx->image_embd_v.data();
 }
 
-size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks) {
-    size_t n_tokens = 0;
-    for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
-        auto chunk = mtmd_input_chunks_get(chunks, i);
-        auto chunk_type = mtmd_input_chunk_get_type(chunk);
-        if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
-            size_t n_tokens_text;
-            mtmd_input_chunk_get_tokens_text(chunk, &n_tokens_text);
-            n_tokens += n_tokens_text;
-        } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-            auto tokens_image = mtmd_input_chunk_get_tokens_image(chunk);
-            n_tokens += mtmd_image_tokens_get_n_tokens(tokens_image);
-        } else {
-            GGML_ASSERT(false && "chunk type not supported");
-        }
+bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
+    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
+    if (proj_type == PROJECTOR_TYPE_GEMMA3) {
+        return true;
     }
-    return n_tokens;
+    return false;
 }
 
-llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks) {
-    llama_pos n_pos = 0;
-    for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
-        auto chunk = mtmd_input_chunks_get(chunks, i);
-        auto chunk_type = mtmd_input_chunk_get_type(chunk);
-        if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
-            size_t n_tokens_text;
-            mtmd_input_chunk_get_tokens_text(chunk, &n_tokens_text);
-            n_pos += n_tokens_text;
-        } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-            auto tokens_image = mtmd_input_chunk_get_tokens_image(chunk);
-            n_pos += mtmd_image_tokens_get_n_pos(tokens_image);
-        } else {
-            GGML_ASSERT(false && "chunk type not supported");
-        }
-    }
-    return n_pos;
+bool mtmd_decode_use_mrope(mtmd_context * ctx) {
+    return ctx->use_mrope;
 }
 
-// helper struct to make working with embd batch easier
-// note: this will be removed after llama_batch_ext refactoring
-struct decode_embd_batch {
-    int n_pos_per_embd;
-    int n_mmproj_embd;
-    std::vector<llama_pos>      pos;
-    std::vector<llama_pos>      pos_view; // used by mrope
-    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, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
-        pos     .resize(n_tokens * n_pos_per_embd);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        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(),
-        };
-    }
-
-    void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {
-        seq_id_0[0] = seq_id;
-        for (int i = 0; i < batch.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;
-        }
-    }
-
-    void set_position_mrope(llama_pos pos_0, int nx, int ny, llama_seq_id seq_id) {
-        GGML_ASSERT(n_pos_per_embd == 4);
-        seq_id_0[0] = seq_id;
-        for (int y = 0; y < ny; y++) {
-            for (int x = 0; x < nx; x++) {
-                int i = y * nx + x;
-                pos[i                     ] = pos_0;
-                pos[i + batch.n_tokens    ] = pos_0 + y;
-                pos[i + batch.n_tokens * 2] = pos_0 + x;
-                pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
-            }
-        }
-        for (int i = 0; i < batch.n_tokens; i++) {
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-
-    llama_batch get_view(int offset, int n_tokens) {
-        llama_pos * pos_ptr;
-        pos_view.clear();
-        pos_view.resize(n_tokens * n_pos_per_embd);
-        if (n_pos_per_embd > 1) {
-            // mrope
-            // for example, with layout of src: 1234...1234...1234...1234...
-            //       offset 2 will give us dst: 34...34...34...34...
-            for (int i = 0; i < n_pos_per_embd; i++) {
-                auto src = pos.begin() + i * batch.n_tokens + offset;
-                pos_view.insert(pos_view.end(), src, src + n_tokens);
-            }
-            pos_ptr = pos_view.data();
-        } else {
-            // normal
-            pos_ptr = pos.data() + offset;
-        }
-        return {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ batch.embd     + offset * n_mmproj_embd,
-            /*pos            =*/ pos_ptr,
-            /*n_seq_id       =*/ batch.n_seq_id + offset,
-            /*seq_id         =*/ batch.seq_id   + offset,
-            /*logits         =*/ batch.logits   + offset,
-        };
-    }
-};
-
-int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
-        struct llama_context * lctx,
-        const mtmd_input_chunk * chunk,
-        llama_pos n_past,
-        llama_seq_id seq_id,
-        int32_t n_batch,
-        bool logits_last,
-        llama_pos * new_n_past) {
-    int32_t ret;
-    llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
-    auto chunk_type = mtmd_input_chunk_get_type(chunk);
-    int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
-    int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;
-
-    if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
-        size_t n_tokens;
-        const auto tokens = mtmd_input_chunk_get_tokens_text(chunk, &n_tokens);
-        LOG_DBG("decoding text chunk, n_tokens = %zu\n", n_tokens);
-        size_t i = 0;
-        while (i < n_tokens) { // split into batches
-            text_batch.n_tokens = 0; // clear the batch
-            for (; i < n_tokens && text_batch.n_tokens < n_batch; i++) {
-                text_batch.n_tokens++;
-                text_batch.token   [i]    = tokens[i];
-                text_batch.pos     [i]    = n_past++;
-                text_batch.n_seq_id[i]    = 1;
-                text_batch.seq_id  [i][0] = seq_id;
-                text_batch.logits  [i]    = false;
-            }
-            bool is_last_token = (i == n_tokens);
-            if (logits_last && is_last_token) {
-                text_batch.logits[text_batch.n_tokens - 1] = true;
-            }
-            ret = llama_decode(lctx, text_batch);
-            if (ret != 0) {
-                LOG_ERR("failed to decode text\n");
-                llama_batch_free(text_batch);
-                return ret;
-            }
-            *new_n_past += text_batch.n_tokens;
-        }
-
-    } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-        const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk);
-        int64_t t0 = ggml_time_ms();
-        if (ctx->print_timings) {
-            LOG_INF("encoding image or slice...\n");
-        }
-        ret = mtmd_encode(ctx, image_tokens);
-        if (ret != 0) {
-            LOG_ERR("failed to encode image\n");
-            llama_batch_free(text_batch);
-            return ret;
-        }
-        if (ctx->print_timings) {
-            LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
-        }
-
-        int32_t n_tokens = mtmd_image_tokens_get_n_tokens(image_tokens);
-        int32_t i_batch = 0;
-        int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
-        float * embd = mtmd_get_output_embd(ctx);
-        decode_embd_batch batch_embd(embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
-
-        const int nx = mtmd_image_tokens_get_nx(image_tokens);
-        const int ny = mtmd_image_tokens_get_ny(image_tokens);
-
-        if (mtmd_decode_use_mrope(ctx)) {
-            batch_embd.set_position_mrope(n_past, nx, ny, seq_id);
-        } else {
-            batch_embd.set_position_normal(n_past, seq_id);
-        }
-
-        if (mtmd_decode_use_non_causal(ctx)) {
-            llama_set_causal_attn(lctx, false);
-            // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
-        }
-
-        while (i_batch < n_img_batches) { // split into batches
-            int pos_offset = i_batch*n_batch;
-            int n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
-            llama_batch batch_embd_view = batch_embd.get_view(pos_offset, n_tokens_batch);
-
-            LOG_INF("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
-
-            int64_t t1 = ggml_time_ms();
-            ret = llama_decode(lctx, batch_embd_view);
-            if (ret != 0) {
-                LOG_ERR("failed to decode image\n");
-                llama_set_causal_attn(lctx, true); // restore causal attn
-                llama_batch_free(text_batch);
-                return ret;
-            }
-
-            if (ctx->print_timings) {
-                LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1);
-            }
-
-            i_batch++;
-        }
-
-        n_past += mtmd_image_tokens_get_n_pos(image_tokens);
-        *new_n_past = n_past;
-
-        if (mtmd_decode_use_non_causal(ctx)) {
-            llama_set_causal_attn(lctx, true);
-        }
-
-    } else {
-        GGML_ABORT("chunk type not supported");
-    }
-
-    return 0;
+void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
+    mtmd_image_tokens_free(val);
 }
 
-int32_t mtmd_helper_eval_chunks(mtmd_context * ctx,
-                                struct llama_context * lctx,
-                                const mtmd_input_chunks * chunks,
-                                llama_pos n_past,
-                                llama_seq_id seq_id,
-                                int32_t n_batch,
-                                bool logits_last,
-                                llama_pos * new_n_past) {
-    size_t n_chunks = mtmd_input_chunks_size(chunks);
-    if (n_chunks == 0) {
-        LOG_WRN("no chunks to eval\n");
-        return 0;
-    }
-
-    for (size_t i = 0; i < n_chunks; i++) {
-        bool chunk_logits_last = (i == n_chunks - 1) && logits_last;
-        auto chunk = mtmd_input_chunks_get(chunks, i);
-
-        int32_t res = mtmd_helper_eval_chunk_single(ctx, lctx, chunk, n_past, seq_id, n_batch, chunk_logits_last, &n_past);
-        if (res != 0) {
-            LOG_ERR("failed to eval chunk %zu\n", i);
-            return res;
-        }
-        *new_n_past = n_past;
-    }
-
-    return 0;
-}
+// these 2 helpers below use internal clip_image_u8_ptr,
+// so unfortunately they cannot moved to mtmd-helper.h
+// however, in theory, user can decode image file to bitmap using
+// whichever library they want, and then use mtmd_bitmap_init() to create bitmap
 
 mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len) {
     clip_image_u8_ptr img_u8(clip_image_u8_init());
@@ -748,23 +506,6 @@ mtmd_bitmap * mtmd_helper_bitmap_init_from_file(const char * fname) {
     return mtmd_bitmap_init(nx, ny, data);
 }
 
-bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
-    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
-    if (proj_type == PROJECTOR_TYPE_GEMMA3) {
-        return true;
-    }
-    return false;
-}
-
-bool mtmd_decode_use_mrope(mtmd_context * ctx) {
-    return ctx->use_mrope;
-}
-
-void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
-    mtmd_image_tokens_free(val);
-}
-
-
 //
 // public API functions
 //

tools/mtmd/mtmd.h

@@ -10,6 +10,7 @@
 #include <stdbool.h>
 
 #ifdef __cplusplus
+#include <string>
 #include <vector>
 #include <cinttypes>
 #include <memory>
@@ -231,6 +232,18 @@ MTMD_API int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
                                                bool logits_last,
                                                llama_pos * new_n_past);
 
+// helper function to decode an image whose embeddings have already been calculated
+// this helper will handle batching and pre/post decoding setup (for ex. gemma 3 requires non-causal attention)
+// ret 0 on success, -1 on chunk not being a valid image chunk, 1 on decode failure
+MTMD_API int32_t mtmd_helper_decode_image_chunk(mtmd_context * ctx,
+                                                struct llama_context * lctx,
+                                                const mtmd_input_chunk * chunk,
+                                                float * encoded_embd,
+                                                llama_pos n_past,
+                                                llama_seq_id seq_id,
+                                                int32_t n_batch,
+                                                llama_pos * new_n_past);
+
 /////////////////////////////////////////
 
 // test function, to be used in test-mtmd-c-api.c

tools/mtmd/tests.sh

@@ -40,7 +40,6 @@ add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-500M-Instruct-GGUF:Q8_0"
 add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-2.2B-Instruct-GGUF:Q4_K_M"
 add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-500M-Video-Instruct-GGUF:Q8_0"
 add_test "llama-mtmd-cli"  "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
-add_test "llama-mtmd-cli"  "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"               "deepseek"
 add_test "llama-mtmd-cli"  "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
 add_test "llama-mtmd-cli"  "second-state/Llava-v1.5-7B-GGUF:Q2_K"            "vicuna"
 add_test "llama-mtmd-cli"  "cjpais/llava-1.6-mistral-7b-gguf:Q3_K"           "vicuna"
@@ -50,6 +49,8 @@ add_test "llama-mtmd-cli"  "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
 add_test "llama-mtmd-cli"  "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
 add_test "llama-mtmd-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
 add_test "llama-mtmd-cli"  "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "ggml-org/InternVL2_5-1B-GGUF:Q8_0"
+add_test "llama-mtmd-cli"  "ggml-org/InternVL3-1B-Instruct-GGUF:Q8_0"
 
 # to test the big models, run: ./tests.sh big
 if [ "$RUN_BIG_TESTS" = true ]; then
@@ -59,6 +60,8 @@ if [ "$RUN_BIG_TESTS" = true ]; then
     add_test "llama-mtmd-cli" "ggml-org/Qwen2-VL-7B-Instruct-GGUF:Q4_K_M"
     add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
     add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-7B-Instruct-GGUF:Q4_K_M"
+    add_test "llama-mtmd-cli"  "ggml-org/InternVL3-8B-Instruct-GGUF:Q4_K_M"
+    add_test "llama-mtmd-cli"  "ggml-org/InternVL3-14B-Instruct-GGUF:Q4_K_M"
     # add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-32B-Instruct-GGUF:Q4_K_M" # does not work on my mac M3 Ultra
     # add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-72B-Instruct-GGUF:Q4_K_M" # too big
 fi
@@ -70,6 +73,7 @@ fi
 
 # this model has broken chat template, not usable
 # add_test "llama-mtmd-cli"  "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
+# add_test "llama-mtmd-cli"  "guinmoon/MobileVLM-3B-GGUF:Q4_K_M" "deepseek"
 
 ###############
 

tools/rpc/rpc-server.cpp

@@ -237,15 +237,17 @@ static ggml_backend_t create_backend(const rpc_server_params & params) {
         backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
     }
 
-    fprintf(stderr, "%s: using %s backend\n", __func__, ggml_backend_name(backend));
+    if (backend) {
+        fprintf(stderr, "%s: using %s backend\n", __func__, ggml_backend_name(backend));
 
-    // set the number of threads
-    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
-    ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
-    if (reg) {
-        auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
-        if (ggml_backend_set_n_threads_fn) {
-            ggml_backend_set_n_threads_fn(backend, params.n_threads);
+        // set the number of threads
+        ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+        ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
+        if (reg) {
+            auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
+            if (ggml_backend_set_n_threads_fn) {
+                ggml_backend_set_n_threads_fn(backend, params.n_threads);
+            }
         }
     }
 

tools/run/README.md

@@ -42,6 +42,8 @@ Examples:
   llama-run ollama://smollm:135m
   llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf
   llama-run huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf
+  llama-run ms://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf
+  llama-run modelscope://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf
   llama-run https://example.com/some-file1.gguf
   llama-run some-file2.gguf
   llama-run file://some-file3.gguf

tools/run/run.cpp

@@ -267,7 +267,7 @@ class Opt {
             "Commands:\n"
             "  model\n"
             "      Model is a string with an optional prefix of \n"
-            "      huggingface:// (hf://), ollama://, https:// or file://.\n"
+            "      huggingface:// (hf://), modelscope:// (ms://), ollama://, https:// or file://.\n"
             "      If no protocol is specified and a file exists in the specified\n"
             "      path, file:// is assumed, otherwise if a file does not exist in\n"
             "      the specified path, ollama:// is assumed. Models that are being\n"
@@ -282,6 +282,9 @@ class Opt {
             "  llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf\n"
             "  llama-run "
             "huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf\n"
+            "  llama-run ms://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf\n"
+            "  llama-run "
+            "modelscope://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf\n"
             "  llama-run https://example.com/some-file1.gguf\n"
             "  llama-run some-file2.gguf\n"
             "  llama-run file://some-file3.gguf\n"
@@ -689,7 +692,7 @@ class LlamaData {
         return 0;
     }
 
-    int huggingface_dl(std::string & model, const std::string & bn) {
+    int dl_from_endpoint(std::string & model_endpoint, std::string & model, const std::string & bn) {
         // Find the second occurrence of '/' after protocol string
         size_t pos = model.find('/');
         pos        = model.find('/', pos + 1);
@@ -697,8 +700,6 @@ class LlamaData {
         std::vector<std::string> headers = { "User-Agent: llama-cpp", "Accept: application/json" };
         std::string              url;
 
-        std::string model_endpoint = get_model_endpoint();
-
         if (pos == std::string::npos) {
             auto [model_name, manifest_url] = extract_model_and_tag(model, model_endpoint + "v2/");
             hfr                             = model_name;
@@ -720,6 +721,16 @@ class LlamaData {
         return download(url, bn, true, headers);
     }
 
+    int modelscope_dl(std::string & model, const std::string & bn) {
+        std::string model_endpoint = "https://modelscope.cn/models/";
+        return dl_from_endpoint(model_endpoint, model, bn);
+    }
+
+    int huggingface_dl(std::string & model, const std::string & bn) {
+        std::string model_endpoint = get_model_endpoint();
+        return dl_from_endpoint(model_endpoint, model, bn);
+    }
+
     int ollama_dl(std::string & model, const std::string & bn) {
         const std::vector<std::string> headers = { "Accept: application/vnd.docker.distribution.manifest.v2+json" };
         if (model.find('/') == std::string::npos) {
@@ -837,6 +848,9 @@ class LlamaData {
             rm_until_substring(model_, "hf.co/");
             rm_until_substring(model_, "://");
             ret = huggingface_dl(model_, bn);
+        } else if (string_starts_with(model_, "ms://") || string_starts_with(model_, "modelscope://")) {
+            rm_until_substring(model_, "://");
+            ret = modelscope_dl(model_, bn);
         } else if ((string_starts_with(model_, "https://") || string_starts_with(model_, "http://")) &&
                    !string_starts_with(model_, "https://ollama.com/library/")) {
             ret = download(model_, bn, true);

tools/server/CMakeLists.txt

@@ -34,8 +34,9 @@ endforeach()
 add_executable(${TARGET} ${TARGET_SRCS})
 install(TARGETS ${TARGET} RUNTIME)
 
+target_include_directories(${TARGET} PRIVATE ../llava)
 target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
-target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
 
 if (LLAMA_SERVER_SSL)
     find_package(OpenSSL REQUIRED)

tools/server/README.md

@@ -7,13 +7,15 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
 **Features:**
  * LLM inference of F16 and quantized models on GPU and CPU
  * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
- * Reranking endoint (WIP: https://github.com/ggml-org/llama.cpp/pull/9510)
+ * Reranking endoint (https://github.com/ggml-org/llama.cpp/pull/9510)
  * Parallel decoding with multi-user support
  * Continuous batching
- * Multimodal (wip)
+ * Multimodal ([documentation](../../docs/multimodal.md)) / with OpenAI-compatible API support
  * Monitoring endpoints
  * Schema-constrained JSON response format
  * [Function calling](../../docs/function-calling.md) / tool use for ~any model
+ * Speculative decoding
+ * Easy-to-use web UI
 
 The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggml-org/llama.cpp/issues/4216).
 
@@ -27,6 +29,7 @@ The project is under active development, and we are [looking for feedback and co
 | -------- | ----------- |
 | `-h, --help, --usage` | print usage and exit |
 | `--version` | show version and build info |
+| `--completion-bash` | print source-able bash completion script for llama.cpp |
 | `--verbose-prompt` | print a verbose prompt before generation (default: false) |
 | `-t, --threads N` | number of threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
 | `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
@@ -41,7 +44,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
 | `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) |
 | `-c, --ctx-size N` | size of the prompt context (default: 4096, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
-| `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)<br/>(env: LLAMA_ARG_N_PREDICT) |
+| `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity)<br/>(env: LLAMA_ARG_N_PREDICT) |
 | `-b, --batch-size N` | logical maximum batch size (default: 2048)<br/>(env: LLAMA_ARG_BATCH) |
 | `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
 | `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
@@ -69,6 +72,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |
+| `--override-tensor, -ot <tensor name pattern>=<buffer type>,...` | override tensor buffer type |
 | `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
 | `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs<br/>(env: LLAMA_ARG_SPLIT_MODE) |
 | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1<br/>(env: LLAMA_ARG_TENSOR_SPLIT) |
@@ -82,15 +86,18 @@ The project is under active development, and we are [looking for feedback and co
 | `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive |
 | `-m, --model FNAME` | model path (default: `models/$filename` with filename from `--hf-file` or `--model-url` if set, otherwise models/7B/ggml-model-f16.gguf)<br/>(env: LLAMA_ARG_MODEL) |
 | `-mu, --model-url MODEL_URL` | model download url (default: unused)<br/>(env: LLAMA_ARG_MODEL_URL) |
-| `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
-| `-hff, --hf-file FILE` | Hugging Face model file (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
+| `-hf, -hfr, --hf-repo <user>/<model>[:quant]` | Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.<br/>mmproj is also downloaded automatically if available. to disable, add --no-mmproj<br/>example: unsloth/phi-4-GGUF:q4_k_m<br/>(default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
+| `-hfd, -hfrd, --hf-repo-draft <user>/<model>[:quant]` | Same as --hf-repo, but for the draft model (default: unused)<br/>(env: LLAMA_ARG_HFD_REPO) |
+| `-hff, --hf-file FILE` | Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
+| `-hfv, -hfrv, --hf-repo-v <user>/<model>[:quant]` | Hugging Face model repository for the vocoder model (default: unused)<br/>(env: LLAMA_ARG_HF_REPO_V) |
+| `-hffv, --hf-file-v FILE` | Hugging Face model file for the vocoder model (default: unused)<br/>(env: LLAMA_ARG_HF_FILE_V) |
 | `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)<br/>(env: HF_TOKEN) |
 | `--log-disable` | Log disable |
 | `--log-file FNAME` | Log to file |
 | `--log-colors` | Enable colored logging<br/>(env: LLAMA_LOG_COLORS) |
 | `-v, --verbose, --log-verbose` | Set verbosity level to infinity (i.e. log all messages, useful for debugging) |
 | `-lv, --verbosity, --log-verbosity N` | Set the verbosity threshold. Messages with a higher verbosity will be ignored.<br/>(env: LLAMA_LOG_VERBOSITY) |
-| `--log-prefix` | Enable prefx in log messages<br/>(env: LLAMA_LOG_PREFIX) |
+| `--log-prefix` | Enable prefix in log messages<br/>(env: LLAMA_LOG_PREFIX) |
 | `--log-timestamps` | Enable timestamps in log messages<br/>(env: LLAMA_LOG_TIMESTAMPS) |
 
 
@@ -98,9 +105,9 @@ The project is under active development, and we are [looking for feedback and co
 
 | Argument | Explanation |
 | -------- | ----------- |
-| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: dry;top_k;typ_p;top_p;min_p;xtc;temperature) |
+| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: penalties;dry;top_n_sigma;top_k;typ_p;top_p;min_p;xtc;temperature) |
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
-| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: dkypmxt) |
+| `--sampling-seq, --sampler-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
 | `--temp N` | temperature (default: 0.8) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
@@ -127,22 +134,26 @@ The project is under active development, and we are [looking for feedback and co
 | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
 | `--grammar-file FNAME` | file to read grammar from |
 | `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
-| `--jinja` | Enable experimental Jinja templating engine (required for tool use) |
-| `--reasoning-format FORMAT` | Controls extraction of model thinking traces and the format / field in which they are returned (default: `deepseek`; allowed values: `deepseek`, `none`; requires `--jinja`). `none` will leave thinking traces inline in `message.content` in a model-specific format, while `deepseek` will return them separately under `message.reasoning_content` |
+| `-jf, --json-schema-file FILE` | File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+
 
 **Example-specific params**
 
 | Argument | Explanation |
 | -------- | ----------- |
-| `--no-context-shift` | disables context shift on inifinite text generation (default: disabled)<br/>(env: LLAMA_ARG_NO_CONTEXT_SHIFT) |
+| `--no-context-shift` | disables context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_NO_CONTEXT_SHIFT) |
 | `-sp, --special` | special tokens output enabled (default: false) |
 | `--no-warmup` | skip warming up the model with an empty run |
 | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
 | `--pooling {none,mean,cls,last,rank}` | pooling type for embeddings, use model default if unspecified<br/>(env: LLAMA_ARG_POOLING) |
 | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
 | `-nocb, --no-cont-batching` | disable continuous batching<br/>(env: LLAMA_ARG_NO_CONT_BATCHING) |
+| `--mmproj FILE` | path to a multimodal projector file. see tools/mtmd/README.md<br/>note: if -hf is used, this argument can be omitted<br/>(env: LLAMA_ARG_MMPROJ) |
+| `--mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md<br/>(env: LLAMA_ARG_MMPROJ_URL) |
+| `--no-mmproj` | explicitly disable multimodal projector, useful when using -hf<br/>(env: LLAMA_ARG_NO_MMPROJ) |
+| `--no-mmproj-offload` | do not offload multimodal projector to GPU<br/>(env: LLAMA_ARG_NO_MMPROJ_OFFLOAD) |
 | `-a, --alias STRING` | set alias for model name (to be used by REST API)<br/>(env: LLAMA_ARG_ALIAS) |
-| `--host HOST` | ip address to listen (default: 127.0.0.1)<br/>(env: LLAMA_ARG_HOST) |
+| `--host HOST` | ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: 127.0.0.1)<br/>(env: LLAMA_ARG_HOST) |
 | `--port PORT` | port to listen (default: 8080)<br/>(env: LLAMA_ARG_PORT) |
 | `--path PATH` | path to serve static files from (default: )<br/>(env: LLAMA_ARG_STATIC_PATH) |
 | `--no-webui` | Disable the Web UI (default: enabled)<br/>(env: LLAMA_ARG_NO_WEBUI) |
@@ -160,16 +171,29 @@ The project is under active development, and we are [looking for feedback and co
 | `--props` | enable changing global properties via POST /props (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_PROPS) |
 | `--no-slots` | disables slots monitoring endpoint<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
 | `--slot-save-path PATH` | path to save slot kv cache (default: disabled) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>list of built-in templates:<br/>chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, exaone3, gemma, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, monarch, openchat, orion, phi3, rwkv-world, vicuna, vicuna-orca, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--jinja` | use jinja template for chat (default: disabled)<br/>(env: LLAMA_ARG_JINJA) |
+| `--reasoning-format FORMAT` | reasoning format (default: deepseek; allowed values: deepseek, none)<br/>controls whether thought tags are extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only).<br/>only supported for non-streamed responses<br/>(env: LLAMA_ARG_THINK) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, falcon3, gemma, gigachat, glmedge, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, falcon3, gemma, gigachat, glmedge, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
 | `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
-| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 5)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
-| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.9)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
+| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
+| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
 | `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE_DRAFT) |
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
 | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
 | `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_MODEL_DRAFT) |
+| `-mv, --model-vocoder FNAME` | vocoder model for audio generation (default: unused) |
+| `--tts-use-guide-tokens` | Use guide tokens to improve TTS word recall |
+| `--embd-bge-small-en-default` | use default bge-small-en-v1.5 model (note: can download weights from the internet) |
+| `--embd-e5-small-en-default` | use default e5-small-v2 model (note: can download weights from the internet) |
+| `--embd-gte-small-default` | use default gte-small model (note: can download weights from the internet) |
+| `--fim-qwen-1.5b-default` | use default Qwen 2.5 Coder 1.5B (note: can download weights from the internet) |
+| `--fim-qwen-3b-default` | use default Qwen 2.5 Coder 3B (note: can download weights from the internet) |
+| `--fim-qwen-7b-default` | use default Qwen 2.5 Coder 7B (note: can download weights from the internet) |
+| `--fim-qwen-7b-spec` | use Qwen 2.5 Coder 7B + 0.5B draft for speculative decoding (note: can download weights from the internet) |
+| `--fim-qwen-14b-spec` | use Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding (note: can download weights from the internet) |
 
 
 Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.
@@ -193,6 +217,12 @@ services:
       LLAMA_ARG_PORT: 8080
 ```
 
+### Multimodal support
+
+Multimodal support was added in [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) and is currently an experimental feature.
+
+For more details, please refer to [multimodal documentation](../../docs/multimodal.md)
+
 ## Build
 
 `llama-server` is built alongside everything else from the root of the project
@@ -749,6 +779,9 @@ This endpoint is public (no API key check). By default, it is read-only. To make
   "total_slots": 1,
   "model_path": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
   "chat_template": "...",
+  "modalities": {
+    "vision": false
+  },
   "build_info": "b(build number)-(build commit hash)"
 }
 ```
@@ -757,6 +790,7 @@ This endpoint is public (no API key check). By default, it is read-only. To make
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 - `model_path` - the path to model file (same with `-m` argument)
 - `chat_template` - the model's original Jinja2 prompt template
+- `modalities` - the list of supported modalities
 
 ### POST `/props`: Change server global properties.
 
@@ -1069,6 +1103,8 @@ print(completion.choices[0].text)
 
 Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
 
+If model supports multimodal, you can input the media file via `image_url` content part. We support both base64 and remote URL as input. See OAI documentation for more.
+
 *Options:*
 
 See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). llama.cpp `/completion`-specific features such as `mirostat` are also supported.

tools/server/server.cpp

@@ -7,6 +7,7 @@
 #include "log.h"
 #include "sampling.h"
 #include "speculative.h"
+#include "mtmd.h"
 
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
 #define JSON_ASSERT GGML_ASSERT
@@ -197,8 +198,8 @@ struct server_task {
     int id_target = -1;
 
     // used by SERVER_TASK_TYPE_INFERENCE
-    slot_params  params;
-    llama_tokens prompt_tokens;
+    slot_params   params;
+    server_tokens prompt_tokens;
     int id_selected_slot = -1;
 
     // used by SERVER_TASK_TYPE_SLOT_SAVE, SERVER_TASK_TYPE_SLOT_RESTORE, SERVER_TASK_TYPE_SLOT_ERASE
@@ -1248,6 +1249,9 @@ struct server_slot {
     llama_context * ctx = nullptr;
     llama_context * ctx_dft = nullptr;
 
+    // multimodal
+    mtmd_context * mctx = nullptr;
+
     common_speculative * spec = nullptr;
 
     std::vector<common_adapter_lora_info> lora;
@@ -1275,14 +1279,14 @@ struct server_slot {
     int32_t n_prompt_tokens_processed = 0;
 
     // input prompt tokens
-    llama_tokens prompt_tokens;
+    server_tokens prompt_tokens;
 
     size_t last_nl_pos = 0;
 
     std::string  generated_text;
     llama_tokens generated_tokens;
 
-    llama_tokens cache_tokens;
+    server_tokens cache_tokens;
 
     std::vector<completion_token_output> generated_token_probs;
 
@@ -1476,7 +1480,7 @@ struct server_slot {
             {"is_processing", is_processing()},
             {"non_causal",    is_non_causal()},
             {"params",        params.to_json()},
-            {"prompt",        common_detokenize(ctx, prompt_tokens)},
+            {"prompt",        prompt_tokens.detokenize(ctx, true)},
             {"next_token",
                 {
                     {"has_next_token", has_next_token},
@@ -1849,13 +1853,16 @@ struct server_context {
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
 
+    // multimodal
+    mtmd_context * mctx = nullptr;
+
     const llama_vocab * vocab = nullptr;
 
     llama_model * model_dft = nullptr;
 
     llama_context_params cparams_dft;
 
-    llama_batch batch = {};
+    llama_batch batch;
 
     bool clean_kv_cache = true;
     bool add_bos_token  = true;
@@ -1878,6 +1885,8 @@ struct server_context {
     common_chat_templates_ptr chat_templates;
 
     ~server_context() {
+        mtmd_free(mctx);
+
         // Clear any sampling context
         for (server_slot & slot : slots) {
             common_sampler_free(slot.smpl);
@@ -1965,6 +1974,36 @@ struct server_context {
             chat_templates = common_chat_templates_init(model, "chatml");
         }
 
+        std::string & mmproj_path = params_base.mmproj.path;
+        if (!mmproj_path.empty()) {
+            mtmd_context_params mparams = mtmd_context_params_default();
+            mparams.use_gpu       = params_base.mmproj_use_gpu;
+            mparams.print_timings = false;
+            mparams.n_threads     = params_base.cpuparams.n_threads;
+            mparams.verbosity     = params_base.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_INFO;
+            mctx = mtmd_init_from_file(mmproj_path.c_str(), model, mparams);
+            if (mctx == nullptr) {
+                SRV_ERR("failed to load multimodal model, '%s'\n", mmproj_path.c_str());
+                return false;
+            }
+            SRV_INF("loaded multimodal model, '%s'\n", mmproj_path.c_str());
+
+            if (params_base.ctx_shift) {
+                params_base.ctx_shift = false;
+                SRV_WRN("%s\n", "ctx_shift is not supported by multimodal, it will be disabled");
+            }
+
+            if (params_base.n_cache_reuse) {
+                params_base.n_cache_reuse = 0;
+                SRV_WRN("%s\n", "cache_reuse is not supported by multimodal, it will be disabled");
+            }
+
+            if (!params_base.speculative.model.path.empty()) {
+                SRV_ERR("%s\n", "err: speculative decode is not supported by multimodal");
+                return false;
+            }
+        }
+
         return true;
     }
 
@@ -1980,6 +2019,8 @@ struct server_context {
             slot.ctx = ctx;
             slot.n_ctx = n_ctx_slot;
             slot.n_predict = params_base.n_predict;
+            slot.mctx = mctx;
+            slot.cache_tokens.has_mtmd = mctx != nullptr;
 
             if (model_dft) {
                 slot.batch_spec = llama_batch_init(params_base.speculative.n_max + 1, 0, 1);
@@ -2016,8 +2057,6 @@ struct server_context {
         // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used)
         {
             const int32_t n_batch = llama_n_batch(ctx);
-
-            // only a single seq_id per token is needed
             batch = llama_batch_init(std::max(n_batch, params_base.n_parallel), 0, 1);
         }
 
@@ -2054,7 +2093,7 @@ struct server_context {
                 }
 
                 // length of the Longest Common Subsequence between the current slot's prompt and the input prompt
-                int cur_lcs_len = common_lcs(slot.cache_tokens, task.prompt_tokens);
+                int cur_lcs_len = slot.cache_tokens.get_common_prefix(task.prompt_tokens);
 
                 // fraction of the common subsequence length compared to the current slot's prompt length
                 float cur_similarity = static_cast<float>(cur_lcs_len) / static_cast<int>(slot.cache_tokens.size());
@@ -2096,18 +2135,6 @@ struct server_context {
         return ret;
     }
 
-    bool can_be_detokenized(const struct llama_context * ctx, const std::vector<llama_token> & tokens) {
-        const llama_model * model = llama_get_model(ctx);
-        const llama_vocab * vocab = llama_model_get_vocab(model);
-        const int32_t n_vocab = llama_vocab_n_tokens(vocab);
-        for (const auto & token : tokens) {
-            if (token < 0 || token >= n_vocab) {
-                return false;
-            }
-        }
-        return true;
-    }
-
     bool launch_slot_with_task(server_slot & slot, server_task && task) {
         slot.reset();
         slot.id_task       = task.id;
@@ -2122,8 +2149,7 @@ struct server_context {
             slot.lora = slot.params.lora;
         }
 
-        bool can_detokenize = can_be_detokenized(ctx, slot.prompt_tokens);
-        if (!can_detokenize) {
+        if (!slot.prompt_tokens.validate(ctx)) {
             send_error(task, "Prompt contains invalid tokens", ERROR_TYPE_INVALID_REQUEST);
             return false;
         }
@@ -2385,6 +2411,15 @@ struct server_context {
         queue_results.send(std::move(res));
     }
 
+    // if multimodal is enabled, send an error and return false
+    bool ensure_no_mtmd(const int id_task) {
+        if (mctx) {
+            send_error(id_task, "This feature is not supported by multimodal", ERROR_TYPE_NOT_SUPPORTED);
+            return false;
+        }
+        return true;
+    }
+
     void send_partial_response(server_slot & slot, const completion_token_output & tkn) {
         auto res = std::make_unique<server_task_result_cmpl_partial>();
 
@@ -2424,7 +2459,7 @@ struct server_context {
         res->content         = std::move(slot.generated_text);
         res->tokens          = std::move(slot.generated_tokens);
         res->timings         = slot.get_timings();
-        res->prompt          = common_detokenize(ctx, slot.prompt_tokens, true);
+        res->prompt          = slot.prompt_tokens.detokenize(ctx, true);
         res->response_fields = std::move(slot.params.response_fields);
 
         res->truncated           = slot.truncated;
@@ -2734,6 +2769,10 @@ struct server_context {
                 } break;
             case SERVER_TASK_TYPE_SLOT_SAVE:
                 {
+                    if (!ensure_no_mtmd(task.id)) {
+                        break;
+                    }
+
                     int id_slot = task.slot_action.slot_id;
                     server_slot * slot = get_slot_by_id(id_slot);
                     if (slot == nullptr) {
@@ -2753,7 +2792,8 @@ struct server_context {
                     std::string filename = task.slot_action.filename;
                     std::string filepath = task.slot_action.filepath;
 
-                    const size_t nwrite = llama_state_seq_save_file(ctx, filepath.c_str(), slot->id, slot->cache_tokens.data(), token_count);
+                    const llama_tokens & tokens = slot->cache_tokens.get_text_tokens();
+                    const size_t nwrite = llama_state_seq_save_file(ctx, filepath.c_str(), slot->id, tokens.data(), token_count);
 
                     const int64_t t_end = ggml_time_us();
                     const double t_save_ms = (t_end - t_start) / 1000.0;
@@ -2770,6 +2810,7 @@ struct server_context {
                 } break;
             case SERVER_TASK_TYPE_SLOT_RESTORE:
                 {
+                    if (!ensure_no_mtmd(task.id)) break;
                     int id_slot = task.slot_action.slot_id;
                     server_slot * slot = get_slot_by_id(id_slot);
                     if (slot == nullptr) {
@@ -2788,15 +2829,18 @@ struct server_context {
                     std::string filename = task.slot_action.filename;
                     std::string filepath = task.slot_action.filepath;
 
-                    slot->cache_tokens.resize(slot->n_ctx);
+                    llama_tokens tokens;
+                    tokens.resize(slot->n_ctx);
                     size_t token_count = 0;
-                    size_t nread = llama_state_seq_load_file(ctx, filepath.c_str(), slot->id, slot->cache_tokens.data(), slot->cache_tokens.size(), &token_count);
+                    size_t nread = llama_state_seq_load_file(ctx, filepath.c_str(), slot->id, tokens.data(), tokens.size(), &token_count);
                     if (nread == 0) {
-                        slot->cache_tokens.resize(0);
+                        slot->cache_tokens.clear(); // KV may already been invalidated?
                         send_error(task, "Unable to restore slot, no available space in KV cache or invalid slot save file", ERROR_TYPE_INVALID_REQUEST);
                         break;
                     }
-                    slot->cache_tokens.resize(token_count);
+                    tokens.resize(token_count);
+                    slot->cache_tokens.clear();
+                    slot->cache_tokens.insert(tokens);
 
                     const int64_t t_end = ggml_time_us();
                     const double t_restore_ms = (t_end - t_start) / 1000.0;
@@ -2813,6 +2857,7 @@ struct server_context {
                 } break;
             case SERVER_TASK_TYPE_SLOT_ERASE:
                 {
+                    if (!ensure_no_mtmd(task.id)) break;
                     int id_slot = task.slot_action.slot_id;
                     server_slot * slot = get_slot_by_id(id_slot);
                     if (slot == nullptr) {
@@ -2844,6 +2889,7 @@ struct server_context {
                     res->id = task.id;
                     queue_results.send(std::move(res));
                 } break;
+
         }
     }
 
@@ -2889,6 +2935,12 @@ struct server_context {
                     continue;
                 }
 
+                if (mctx) {
+                    // we should never reach this because params_base.ctx_shift is automatically disabled if mmproj is loaded
+                    // we don't support ctx_shift because an image chunk may contains multiple tokens
+                    GGML_ABORT("not supported by multimodal");
+                }
+
                 // Shift context
                 const int n_keep    = slot.params.n_keep + add_bos_token;
                 const int n_left    = slot.n_past - n_keep;
@@ -2900,11 +2952,14 @@ struct server_context {
                 llama_kv_self_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
 
                 if (slot.params.cache_prompt) {
-                    for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) {
-                        slot.cache_tokens[i - n_discard] = slot.cache_tokens[i];
+                    llama_tokens new_tokens = slot.cache_tokens.get_text_tokens(); // copy
+                    for (size_t i = n_keep + n_discard; i < new_tokens.size(); i++) {
+                        new_tokens[i - n_discard] = new_tokens[i];
                     }
 
-                    slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard);
+                    new_tokens.resize(slot.cache_tokens.size() - n_discard);
+                    slot.cache_tokens.clear();
+                    slot.cache_tokens.insert(new_tokens);
                 }
 
                 slot.n_past -= n_discard;
@@ -2982,7 +3037,7 @@ struct server_context {
                         SLT_INF(slot, "new prompt, n_ctx_slot = %d, n_keep = %d, n_prompt_tokens = %d\n", slot.n_ctx, slot.params.n_keep, slot.n_prompt_tokens);
 
                         // print prompt tokens (for debugging)
-                        if (1) {
+                        /*if (1) {
                             // first 16 tokens (avoid flooding logs)
                             for (int i = 0; i < std::min<int>(16, prompt_tokens.size()); i++) {
                                 SLT_DBG(slot, "prompt token %3d: %6d '%s'\n", i, prompt_tokens[i], common_token_to_piece(ctx, prompt_tokens[i]).c_str());
@@ -2992,7 +3047,7 @@ struct server_context {
                             for (int i = 0; i < (int) prompt_tokens.size(); i++) {
                                 SLT_DBG(slot, "prompt token %3d: %6d '%s'\n", i, prompt_tokens[i], common_token_to_piece(ctx, prompt_tokens[i]).c_str());
                             }
-                        }
+                        }*/
 
                         // empty prompt passed -> release the slot and send empty response
                         if (prompt_tokens.empty()) {
@@ -3034,21 +3089,27 @@ struct server_context {
 
                             // if input prompt is too big, truncate it
                             if (slot.n_prompt_tokens >= slot.n_ctx) {
+                                if (mctx) {
+                                    // we should never reach this
+                                    GGML_ABORT("not supported by multimodal");
+                                }
                                 const int n_left = slot.n_ctx - slot.params.n_keep;
 
                                 const int n_block_size = n_left / 2;
                                 const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size;
 
+                                const llama_tokens & curr_tokens = slot.prompt_tokens.get_text_tokens();
                                 llama_tokens new_tokens(
-                                        prompt_tokens.begin(),
-                                        prompt_tokens.begin() + slot.params.n_keep);
+                                        curr_tokens.begin(),
+                                        curr_tokens.begin() + slot.params.n_keep);
 
                                 new_tokens.insert(
                                         new_tokens.end(),
-                                        prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size,
-                                        prompt_tokens.end());
+                                        curr_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size,
+                                        curr_tokens.end());
 
-                                prompt_tokens = std::move(new_tokens);
+                                prompt_tokens.clear();
+                                prompt_tokens.insert(new_tokens);
 
                                 slot.truncated = true;
                                 slot.n_prompt_tokens = prompt_tokens.size();
@@ -3060,13 +3121,18 @@ struct server_context {
 
                             if (slot.params.cache_prompt) {
                                 // reuse any previously computed tokens that are common with the new prompt
-                                slot.n_past = common_lcp(slot.cache_tokens, prompt_tokens);
+                                slot.n_past = slot.cache_tokens.get_common_prefix(prompt_tokens);
 
                                 // reuse chunks from the cached prompt by shifting their KV cache in the new position
                                 if (params_base.n_cache_reuse > 0) {
                                     size_t head_c = slot.n_past; // cache
                                     size_t head_p = slot.n_past; // current prompt
 
+                                    if (mctx) {
+                                        // we should never reach this
+                                        GGML_ABORT("not supported by multimodal");
+                                    }
+
                                     SLT_DBG(slot, "trying to reuse chunks with size > %d, slot.n_past = %d\n", params_base.n_cache_reuse, slot.n_past);
 
                                     while (head_c < slot.cache_tokens.size() &&
@@ -3092,7 +3158,7 @@ struct server_context {
                                             llama_kv_self_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift);
 
                                             for (size_t i = 0; i < n_match; i++) {
-                                                slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i];
+                                                slot.cache_tokens.set_token(head_p + i, slot.cache_tokens[head_c + i]);
                                                 slot.n_past++;
                                             }
 
@@ -3140,21 +3206,52 @@ struct server_context {
                     // remove the non-common part from the cache
                     slot.cache_tokens.resize(slot.n_past);
 
+                    // check if we should process the image
+                    if (slot.n_past < slot.n_prompt_tokens
+                            && slot.prompt_tokens[slot.n_past] == LLAMA_TOKEN_NULL) {
+                        // process the image
+                        int32_t new_n_past;
+                        int32_t res = slot.prompt_tokens.process_chunk(ctx, mctx, slot.n_past, slot.id, new_n_past);
+                        int32_t n_pos = new_n_past - slot.n_past;
+
+                        if (res != 0) {
+                            SLT_ERR(slot, "failed to process image, res = %d\n", res);
+                            slot.release();
+                            send_error(slot, "failed to process image", ERROR_TYPE_SERVER);
+                            continue;
+                        }
+
+                        if (slot.params.cache_prompt) {
+                            const auto & chunk = slot.prompt_tokens.find_chunk(slot.n_past);
+                            slot.cache_tokens.push_back(chunk.get()); // copy
+                        }
+
+                        slot.n_past                    += n_pos;
+                        slot.n_prompt_tokens_processed += n_pos;
+                    }
+
                     // add prompt tokens for processing in the current batch
                     while (slot.n_past < slot.n_prompt_tokens && batch.n_tokens < n_batch) {
+                        // get next token to process
+                        llama_token cur_tok = slot.prompt_tokens[slot.n_past];
+                        if (cur_tok == LLAMA_TOKEN_NULL) {
+                            break; // end of text chunk
+                        }
+
                         // without pooling, we want to output the embeddings for all the tokens in the batch
                         const bool need_embd = slot.task_type == SERVER_TASK_TYPE_EMBEDDING && llama_pooling_type(slot.ctx) == LLAMA_POOLING_TYPE_NONE;
 
-                        common_batch_add(batch, prompt_tokens[slot.n_past], slot.n_past, { slot.id }, need_embd);
-
+                        common_batch_add(batch, cur_tok, slot.n_past, { slot.id }, need_embd);
                         if (slot.params.cache_prompt) {
-                            slot.cache_tokens.push_back(prompt_tokens[slot.n_past]);
+                            slot.cache_tokens.push_back(cur_tok);
                         }
 
                         slot.n_prompt_tokens_processed++;
                         slot.n_past++;
                     }
 
+                    // SLT_INF(slot, "new cache_tokens: %s\n", slot.cache_tokens.str().c_str());
+
                     SLT_INF(slot, "prompt processing progress, n_past = %d, n_tokens = %d, progress = %f\n", slot.n_past, batch.n_tokens, (float) slot.n_prompt_tokens_processed / slot.n_prompt_tokens);
 
                     // entire prompt has been processed
@@ -3162,12 +3259,16 @@ struct server_context {
                         slot.state = SLOT_STATE_DONE_PROMPT;
 
                         GGML_ASSERT(batch.n_tokens > 0);
+                        GGML_ASSERT((size_t) slot.n_prompt_tokens == slot.prompt_tokens.size());
 
                         common_sampler_reset(slot.smpl);
 
                         // Process all prompt tokens through sampler system
                         for (int i = 0; i < slot.n_prompt_tokens; ++i) {
-                            common_sampler_accept(slot.smpl, prompt_tokens[i], false);
+                            llama_token id = slot.prompt_tokens[i];
+                            if (id != LLAMA_TOKEN_NULL) {
+                                common_sampler_accept(slot.smpl, id, false);
+                            }
                         }
 
                         // extract the logits only for the last token
@@ -3320,6 +3421,11 @@ struct server_context {
                     continue;
                 }
 
+                if (mctx) {
+                    // we should never reach this, as speculative is automatically disabled if mmproj is loaded
+                    GGML_ABORT("not supported by multimodal");
+                }
+
                 // determine the max draft that fits the current slot state
                 int n_draft_max = slot.params.speculative.n_max;
 
@@ -3346,7 +3452,8 @@ struct server_context {
                 params_spec.n_reuse   = llama_n_ctx(slot.ctx_dft) - slot.params.speculative.n_max;
                 params_spec.p_min     = slot.params.speculative.p_min;
 
-                llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
+                const llama_tokens & cached_text_tokens = slot.cache_tokens.get_text_tokens();
+                llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, cached_text_tokens, id);
 
                 // keep track of total number of tokens generated in the draft
                 slot.n_draft_total += draft.size();
@@ -3380,7 +3487,7 @@ struct server_context {
                 slot.n_draft_accepted += ids.size() - 1;
 
                 slot.cache_tokens.push_back(id);
-                slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
+                slot.cache_tokens.insert({ids.begin(), ids.end() - 1});
 
                 llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1);
 
@@ -3903,6 +4010,7 @@ int main(int argc, char ** argv) {
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
             { "total_slots",                 ctx_server.params_base.n_parallel },
             { "model_path",                  ctx_server.params_base.model.path },
+            { "modalities",                  json{{"vision", ctx_server.mctx != nullptr}} }, // TODO: add more in the future
             { "chat_template",               common_chat_templates_source(ctx_server.chat_templates.get()) },
             { "bos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)},
             { "eos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)},
@@ -3950,9 +4058,10 @@ int main(int argc, char ** argv) {
     const auto handle_completions_impl = [&ctx_server, &res_error, &res_ok](
             server_task_type type,
             json & data,
+            const std::vector<raw_buffer> & files,
             const std::function<bool()> & is_connection_closed,
             httplib::Response & res,
-            oaicompat_type oaicompat) {
+            oaicompat_type oaicompat) -> void {
         GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL);
 
         if (ctx_server.params_base.embedding) {
@@ -3969,15 +4078,69 @@ int main(int argc, char ** argv) {
             // TODO: this log can become very long, put it behind a flag or think about a more compact format
             //SRV_DBG("Prompt: %s\n", prompt.is_string() ? prompt.get<std::string>().c_str() : prompt.dump(2).c_str());
 
-            std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
-            tasks.reserve(tokenized_prompts.size());
-            for (size_t i = 0; i < tokenized_prompts.size(); i++) {
+            // process files
+            mtmd::bitmaps bitmaps;
+            const bool has_mtmd = ctx_server.mctx != nullptr;
+            {
+                if (!has_mtmd && !files.empty()) {
+                    throw std::runtime_error("This server does not support multimodal");
+                }
+                for (auto & file : files) {
+                    mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_buf(file.data(), file.size()));
+                    if (!bmp.ptr) {
+                        throw std::runtime_error("Failed to load image");
+                    }
+                    // calculate bitmap hash (for KV caching)
+                    std::string hash = fnv_hash(bmp.data(), bmp.nx()*bmp.ny()*3);
+                    bmp.set_id(hash.c_str());
+                    bitmaps.entries.push_back(std::move(bmp));
+                }
+            }
+
+            // process prompt
+            std::vector<server_tokens> inputs;
+            if (oaicompat && !prompt.is_string()) {
+                throw std::runtime_error("prompt must be a string");
+            }
+
+            if (oaicompat && has_mtmd) {
+                // multimodal
+                std::string prompt_str = prompt.get<std::string>();
+                mtmd_input_text inp_txt = {
+                    prompt_str.c_str(),
+                    /* add_special */   true,
+                    /* parse_special */ true,
+                };
+                mtmd::input_chunks chunks(mtmd_input_chunks_init());
+                auto bitmaps_c_ptr = bitmaps.c_ptr();
+                int32_t tokenized = mtmd_tokenize(ctx_server.mctx,
+                                                    chunks.ptr.get(),
+                                                    &inp_txt,
+                                                    bitmaps_c_ptr.data(),
+                                                    bitmaps_c_ptr.size());
+                if (tokenized != 0) {
+                    throw std::runtime_error("Failed to tokenize prompt");
+                }
+
+                server_tokens tmp(chunks, true);
+                inputs.push_back(std::move(tmp));
+            } else {
+                // non-multimodal version
+                auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
+                for (auto & p : tokenized_prompts) {
+                    auto tmp = server_tokens(p, ctx_server.mctx != nullptr);
+                    inputs.push_back(std::move(tmp));
+                }
+            }
+
+            tasks.reserve(inputs.size());
+            for (size_t i = 0; i < inputs.size(); i++) {
                 server_task task = server_task(type);
 
                 task.id    = ctx_server.queue_tasks.get_new_id();
                 task.index = i;
 
-                task.prompt_tokens    = std::move(tokenized_prompts[i]);
+                task.prompt_tokens    = std::move(inputs[i]);
                 task.params           = server_task::params_from_json_cmpl(
                         ctx_server.ctx,
                         ctx_server.params_base,
@@ -4059,9 +4222,11 @@ int main(int argc, char ** argv) {
 
     const auto handle_completions = [&handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
         json data = json::parse(req.body);
-        return handle_completions_impl(
+        std::vector<raw_buffer> files; // dummy
+        handle_completions_impl(
             SERVER_TASK_TYPE_COMPLETION,
             data,
+            files,
             req.is_connection_closed,
             res,
             OAICOMPAT_TYPE_NONE);
@@ -4069,9 +4234,11 @@ int main(int argc, char ** argv) {
 
     const auto handle_completions_oai = [&handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
         json data = oaicompat_completion_params_parse(json::parse(req.body));
-        return handle_completions_impl(
+        std::vector<raw_buffer> files; // dummy
+        handle_completions_impl(
             SERVER_TASK_TYPE_COMPLETION,
             data,
+            files,
             req.is_connection_closed,
             res,
             OAICOMPAT_TYPE_COMPLETION);
@@ -4146,9 +4313,11 @@ int main(int argc, char ** argv) {
             tokenized_prompts[0]
         );
 
-        return handle_completions_impl(
+        std::vector<raw_buffer> files; // dummy
+        handle_completions_impl(
             SERVER_TASK_TYPE_INFILL,
             data,
+            files,
             req.is_connection_closed,
             res,
             OAICOMPAT_TYPE_NONE); // infill is not OAI compatible
@@ -4162,11 +4331,19 @@ int main(int argc, char ** argv) {
         }
 
         auto body = json::parse(req.body);
-        json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates.get());
+        std::vector<raw_buffer> files;
+        json data = oaicompat_completion_params_parse(
+            body,
+            params.use_jinja,
+            params.reasoning_format,
+            ctx_server.chat_templates.get(),
+            ctx_server.mctx,
+            files);
 
-        return handle_completions_impl(
+        handle_completions_impl(
             SERVER_TASK_TYPE_COMPLETION,
             data,
+            files,
             req.is_connection_closed,
             res,
             OAICOMPAT_TYPE_CHAT);
@@ -4175,7 +4352,14 @@ int main(int argc, char ** argv) {
     // same with handle_chat_completions, but without inference part
     const auto handle_apply_template = [&ctx_server, &params, &res_ok](const httplib::Request & req, httplib::Response & res) {
         auto body = json::parse(req.body);
-        json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates.get());
+        std::vector<raw_buffer> files; // dummy, unused
+        json data = oaicompat_completion_params_parse(
+            body,
+            params.use_jinja,
+            params.reasoning_format,
+            ctx_server.chat_templates.get(),
+            ctx_server.mctx,
+            files);
         res_ok(res, {{ "prompt", std::move(data.at("prompt")) }});
     };
 
@@ -4280,7 +4464,7 @@ int main(int argc, char ** argv) {
             }
         }
 
-        std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
+        auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
         for (const auto & tokens : tokenized_prompts) {
             // this check is necessary for models that do not add BOS token to the input
             if (tokens.empty()) {
@@ -4300,7 +4484,7 @@ int main(int argc, char ** argv) {
 
                 task.id            = ctx_server.queue_tasks.get_new_id();
                 task.index         = i;
-                task.prompt_tokens = std::move(tokenized_prompts[i]);
+                task.prompt_tokens = server_tokens(tokenized_prompts[i], ctx_server.mctx != nullptr);
 
                 // OAI-compat
                 task.params.oaicompat = oaicompat;
@@ -4394,13 +4578,14 @@ int main(int argc, char ** argv) {
         std::unordered_set<int> task_ids;
         {
             std::vector<server_task> tasks;
-            std::vector<llama_tokens> tokenized_docs = tokenize_input_prompts(ctx_server.vocab, documents, /* add_special */ false, true);
+            auto tokenized_docs = tokenize_input_prompts(ctx_server.vocab, documents, /* add_special */ false, true);
             tasks.reserve(tokenized_docs.size());
             for (size_t i = 0; i < tokenized_docs.size(); i++) {
+                auto tmp = format_rerank(ctx_server.vocab, tokenized_query, tokenized_docs[i]);
                 server_task task   = server_task(SERVER_TASK_TYPE_RERANK);
                 task.id            = ctx_server.queue_tasks.get_new_id();
                 task.index         = i;
-                task.prompt_tokens = format_rerank(ctx_server.vocab, tokenized_query, tokenized_docs[i]);
+                task.prompt_tokens = server_tokens(tmp, ctx_server.mctx != nullptr);
                 tasks.push_back(std::move(task));
             }