vulkan: remove the need for the dryrun (#16826)

* vulkan: remove the need for the dryrun

Allocate pipelines and descriptor sets when requested.

Reallocate the prealloc buffers when needed, and flush any pending work
before reallocating.

For rms_partials and total_mul_mat_bytes, use the sizes computed the last time
the graph was executed.

* remove dryrun parameters

.devops/s390x.Dockerfile

@@ -24,8 +24,9 @@ RUN --mount=type=cache,target=/root/.ccache \
         -DCMAKE_C_COMPILER_LAUNCHER=ccache \
         -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
         -DLLAMA_BUILD_TESTS=OFF \
-        -DGGML_BACKEND_DL=OFF \
         -DGGML_NATIVE=OFF \
+        -DGGML_BACKEND_DL=ON \
+        -DGGML_CPU_ALL_VARIANTS=ON \
         -DGGML_BLAS=ON \
         -DGGML_BLAS_VENDOR=OpenBLAS && \
     cmake --build build --config Release -j $(nproc) && \
@@ -103,6 +104,7 @@ FROM base AS light
 WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
+COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
 COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin
 
 ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]
@@ -116,6 +118,7 @@ ENV LLAMA_ARG_HOST=0.0.0.0
 WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
+COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
 COPY --from=collector /llama.cpp/bin/llama-server /llama.cpp/bin
 
 EXPOSE 8080

.github/labeler.yml

@@ -76,6 +76,10 @@ ggml:
     - changed-files:
         - any-glob-to-any-file:
             - ggml/**
+model:
+    - changed-files:
+        - any-glob-to-any-file:
+            - src/models/**
 nix:
     - changed-files:
         - any-glob-to-any-file:

.github/workflows/build-linux-cross.yml

@@ -4,49 +4,49 @@ on:
   workflow_call:
 
 jobs:
-  ubuntu-24-riscv64-cpu-cross:
-    runs-on: ubuntu-24.04
+  # ubuntu-24-riscv64-cpu-cross:
+  #   runs-on: ubuntu-24.04
 
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup Riscv
-        run: |
-          sudo dpkg --add-architecture riscv64
+  #   steps:
+  #     - uses: actions/checkout@v4
+  #     - name: Setup Riscv
+  #       run: |
+  #         sudo dpkg --add-architecture riscv64
 
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
+  #         # Add arch-specific repositories for non-amd64 architectures
+  #         cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+  #         EOF
 
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
 
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  gcc-14-riscv64-linux-gnu \
-                  g++-14-riscv64-linux-gnu
+  #         sudo apt-get install -y --no-install-recommends \
+  #                 build-essential \
+  #                 gcc-14-riscv64-linux-gnu \
+  #                 g++-14-riscv64-linux-gnu
 
-      - name: Build
-        run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+  #     - name: Build
+  #       run: |
+  #         cmake -B build -DLLAMA_CURL=OFF \
+  #                        -DCMAKE_BUILD_TYPE=Release \
+  #                        -DGGML_OPENMP=OFF \
+  #                        -DLLAMA_BUILD_EXAMPLES=ON \
+  #                        -DLLAMA_BUILD_TOOLS=ON \
+  #                        -DLLAMA_BUILD_TESTS=OFF \
+  #                        -DCMAKE_SYSTEM_NAME=Linux \
+  #                        -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+  #                        -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+  #                        -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
-          cmake --build build --config Release -j $(nproc)
+  #         cmake --build build --config Release -j $(nproc)
 
   # ubuntu-24-riscv64-vulkan-cross:
   #   runs-on: ubuntu-24.04

.github/workflows/release.yml

@@ -134,8 +134,8 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
-          - build: 's390x-z15' # z15 because our CI runners are on z15
-            os: ubuntu-22.04-s390x
+          - build: 's390x'
+            os: ubuntu-24.04-s390x
           # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
           # - build: 'arm64'
           #   os: ubuntu-22.04-arm

CODEOWNERS

@@ -89,6 +89,7 @@
 /src/llama-model-loader.*               @slaren
 /src/llama-model.*                      @CISC
 /src/llama-vocab.*                      @CISC
+/src/models/                            @CISC
 /tests/                                 @ggerganov
 /tests/test-backend-ops.cpp             @slaren
 /tests/test-thread-safety.cpp           @slaren

README.md

@@ -17,14 +17,13 @@ LLM inference in C/C++
 
 ## Hot topics
 
-- **[guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)**
-- **[[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)**
+- **[guide : using the new WebUI of llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/16938)**
+- [guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)
+- [[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)
 - Support for the `gpt-oss` model with native MXFP4 format has been added | [PR](https://github.com/ggml-org/llama.cpp/pull/15091) | [Collaboration with NVIDIA](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss) | [Comment](https://github.com/ggml-org/llama.cpp/discussions/15095)
-- Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
 - Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
-- Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
 - Hugging Face GGUF editor: [discussion](https://github.com/ggml-org/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 

common/arg.cpp

@@ -2030,7 +2030,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.system_prompt.pop_back();
             }
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_DIFFUSION}));
     add_opt(common_arg(
         {"--in-file"}, "FNAME",
         "an input file (repeat to specify multiple files)",
@@ -2768,6 +2768,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.image.emplace_back(value);
         }
     ).set_examples({LLAMA_EXAMPLE_MTMD}));
+    add_opt(common_arg(
+        {"--image-min-tokens"}, "N",
+        "minimum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)",
+        [](common_params & params, int value) {
+            params.image_min_tokens = value;
+        }
+    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_IMAGE_MIN_TOKENS"));
+    add_opt(common_arg(
+        {"--image-max-tokens"}, "N",
+        "maximum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)",
+        [](common_params & params, int value) {
+            params.image_max_tokens = value;
+        }
+    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_IMAGE_MAX_TOKENS"));
     if (llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",

common/chat.cpp

@@ -313,7 +313,6 @@ json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msg
         }
         if (!msg.reasoning_content.empty()) {
             jmsg["reasoning_content"] = msg.reasoning_content;
-            jmsg["thinking"] = msg.reasoning_content; // gpt-oss
         }
         if (!msg.tool_name.empty()) {
             jmsg["name"] = msg.tool_name;
@@ -1810,7 +1809,23 @@ static void common_chat_parse_deepseek_v3_1(common_chat_msg_parser & builder) {
 
 static common_chat_params common_chat_params_init_gpt_oss(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    auto prompt = apply(tmpl, inputs);
+
+    // Copy reasoning to the "thinking" field as expected by the gpt-oss template
+    auto adjusted_messages = json::array();
+    for (const auto & msg : inputs.messages) {
+        auto has_reasoning_content = msg.contains("reasoning_content") && msg.at("reasoning_content").is_string();
+        auto has_tool_calls = msg.contains("tool_calls") && msg.at("tool_calls").is_array();
+
+        if (has_reasoning_content && has_tool_calls) {
+            auto adjusted_message = msg;
+            adjusted_message["thinking"] = msg.at("reasoning_content");
+            adjusted_messages.push_back(adjusted_message);
+        } else {
+            adjusted_messages.push_back(msg);
+        }
+    }
+
+    auto prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
 
     // Check if we need to replace the return token with end token during
     // inference and without generation prompt. For more details see:

common/common.h

@@ -406,6 +406,8 @@ struct common_params {
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
+    int image_min_tokens = -1;
+    int image_max_tokens = -1;
 
     // finetune
     struct lr_opt lr;

convert_hf_to_gguf.py

@@ -1054,6 +1054,9 @@ class TextModel(ModelBase):
         if chkhsh == "53e325976a6e142379c19b09afcae354f2f496f147afa8f9e189a33fe4e3024e":
             # ref: https://huggingface.co/ibm-granite/granite-docling-258M
             res = "granite-docling"
+        if chkhsh == "f4f37b6c8eb9ea29b3eac6bb8c8487c5ab7885f8d8022e67edc1c68ce8403e95":
+            # ref: https://huggingface.co/MiniMaxAI/MiniMax-M2
+            res = "minimax-m2"
 
         if res is None:
             logger.warning("\n")
@@ -7126,6 +7129,64 @@ class DeepseekV2Model(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("MiniMaxM2ForCausalLM")
+class MiniMaxM2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.MINIMAXM2
+    _experts_cache: dict[int, dict[str, Tensor]] = {}
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["num_experts"] = self.hparams["num_local_experts"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if self.hparams["scoring_func"] == "sigmoid":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        elif self.hparams["scoring_func"] == "softmax":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
+        else:
+            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
+
+        self.gguf_writer.add_expert_feed_forward_length(self.find_hparam(["intermediate_size"]))
+        self.gguf_writer.add_rope_dimension_count(self.find_hparam(["rotary_dim"]))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        # merge expert weights
+        if 'experts' in name:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            expert_cache = self._experts_cache.setdefault(bid, {})
+            expert_cache[name] = data_torch
+            expert_weights = ["w1", "w2", "w3"]
+
+            # not enough expert weights to merge
+            if len(expert_cache) < n_experts * len(expert_weights):
+                return []
+
+            tensors: list[tuple[str, Tensor]] = []
+            for w_name in expert_weights:
+                datas: list[Tensor] = []
+
+                for xid in range(n_experts):
+                    ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{w_name}.weight"
+                    datas.append(expert_cache[ename])
+                    del expert_cache[ename]
+
+                data_torch = torch.stack(datas, dim=0)
+                merged_name = f"model.layers.{bid}.block_sparse_moe.experts.{w_name}.weight"
+                new_name = self.map_tensor_name(merged_name)
+                tensors.append((new_name, data_torch))
+
+            del self._experts_cache[bid]
+            return tensors
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("Dots1ForCausalLM")
 class Dots1Model(Qwen2MoeModel):
     model_arch = gguf.MODEL_ARCH.DOTS1
@@ -9741,6 +9802,113 @@ class CogVLMModel(LlamaModel):
 
         return [(self.map_tensor_name(name), data_torch)]
 
+
+@ModelBase.register("JanusForConditionalGeneration")
+class JanusProModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA  # reuse Llama arch
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision, aligner, and generation tensors
+        skip_prefixes = (
+            'model.vision_model.',
+            'model.aligner.',
+            'model.vqmodel.',
+            'model.generation_embeddings.',
+            'model.generation_aligner.',
+            'model.generation_head.',
+        )
+        if name.startswith(skip_prefixes):
+            return []
+
+        if name.startswith('model.language_model.'):
+            name = name.replace('model.language_model.', 'model.')
+        elif name.startswith('language_model.'):
+            name = name.replace('language_model.', '')
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("JanusForConditionalGeneration")
+class JanusProVisionModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        if "intermediate_size" not in self.hparams_vision:
+            mlp_ratio = self.hparams_vision.get("mlp_ratio")
+            hidden_size = self.hparams_vision.get("hidden_size")
+            if mlp_ratio is not None and hidden_size is not None:
+                self.hparams_vision["intermediate_size"] = int(round(hidden_size * mlp_ratio))
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.JANUS_PRO)
+
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))
+
+        hidden_act = str(self.hparams_vision.get("hidden_act", "")).lower()
+        if hidden_act == "gelu":
+            self.gguf_writer.add_vision_use_gelu(True)
+        elif hidden_act == "silu":
+            self.gguf_writer.add_vision_use_silu(True)
+
+    def _map_aligner_tensor(self, data_torch: Tensor, name: str) -> Iterable[tuple[str, Tensor]]:
+        """Map aligner tensors to projector format"""
+        suffix = ".bias" if name.endswith(".bias") else ".weight"
+
+        if name.startswith("model.aligner."):
+            local_name = name[len("model.aligner."):]
+        elif name.startswith("aligner."):
+            local_name = name[len("aligner."):]
+        else:
+            raise ValueError(f"Unsupported Janus aligner prefix: {name}")
+
+        if local_name.startswith("fc1."):
+            mm_index = 0
+        elif local_name.startswith("hidden_layers."):
+            parts = local_name.split(".", 2)
+            if len(parts) < 3:
+                raise ValueError(f"Unexpected Janus aligner tensor name: {name}")
+            mm_index = int(parts[1]) + 1
+        else:
+            raise ValueError(f"Unsupported Janus aligner tensor: {name}")
+
+        tensor_name = self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, mm_index, suffix=suffix)
+        return [(tensor_name, data_torch)]
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # Skip language model tensors as they will be handled by `JanusProModel`
+        if name.startswith(('model.language_model.', 'language_model.')):
+            return []
+
+        # Skip generation-related components
+        skip_generation_prefixes = (
+            'model.vqmodel.',
+            'vqmodel.',
+            'model.generation_embeddings.',
+            'generation_embeddings.',
+            'model.generation_aligner.',
+            'generation_aligner.',
+            'model.generation_head.',
+            'generation_head.',
+        )
+        if name.startswith(skip_generation_prefixes):
+            return []
+
+        # Handle aligner tensors
+        if name.startswith(('model.aligner.', 'aligner.')):
+            return list(self._map_aligner_tensor(data_torch, name))
+
+        # Handle vision tensors
+        if name.startswith(('model.vision_model.', 'vision_model.')):
+            return [(self.map_tensor_name(name), data_torch)]
+
+        return []
+
+
 ###### CONVERSION LOGIC ######
 
 

convert_hf_to_gguf_update.py

@@ -141,6 +141,7 @@ models = [
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
     {"name": "bailingmoe2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-mini-base-2.0", },
     {"name": "granite-docling",  "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },
+    {"name": "minimax-m2",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/MiniMaxAI/MiniMax-M2", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -435,7 +436,7 @@ for model in models:
             tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
         else:
             tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-    except OSError as e:
+    except (OSError, TypeError) as e:
         logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
         continue  # Skip this model and continue with the next one in the loop
 

docs/docker.md

@@ -7,9 +7,9 @@
 ## Images
 We have three Docker images available for this project:
 
-1. `ghcr.io/ggml-org/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
-2. `ghcr.io/ggml-org/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggml-org/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
+1. `ghcr.io/ggml-org/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
+2. `ghcr.io/ggml-org/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
+3. `ghcr.io/ggml-org/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
 
 Additionally, there the following images, similar to the above:
 

examples/model-conversion/scripts/causal/run-org-model.py

@@ -138,6 +138,9 @@ if model_path is None:
         "Model path must be specified either via --model-path argument or MODEL_PATH environment variable"
     )
 
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
 
 print("Model type:       ", config.model_type)
@@ -147,10 +150,6 @@ print("Number of layers: ", config.num_hidden_layers)
 print("BOS token id:     ", config.bos_token_id)
 print("EOS token id:     ", config.eos_token_id)
 
-print("Loading model and tokenizer using AutoTokenizer:", model_path)
-tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
-config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
-
 if unreleased_model_name:
     model_name_lower = unreleased_model_name.lower()
     unreleased_module_path = (
@@ -171,7 +170,7 @@ if unreleased_model_name:
         exit(1)
 else:
     model = AutoModelForCausalLM.from_pretrained(
-        model_path, device_map="auto", offload_folder="offload", trust_remote_code=True
+        model_path, device_map="auto", offload_folder="offload", trust_remote_code=True, config=config
     )
 
 for name, module in model.named_modules():

ggml/include/ggml.h

@@ -2108,6 +2108,7 @@ extern "C" {
     enum ggml_scale_mode {
         GGML_SCALE_MODE_NEAREST  = 0,
         GGML_SCALE_MODE_BILINEAR = 1,
+        GGML_SCALE_MODE_BICUBIC  = 2,
 
         GGML_SCALE_MODE_COUNT
     };

ggml/src/CMakeLists.txt

@@ -308,6 +308,10 @@ function(ggml_add_cpu_backend_variant tag_name)
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
+        foreach (feat VXE2 NNPA)
+            set(GGML_INTERNAL_${feat} OFF)
+        endforeach()
+
         foreach (feat ${ARGN})
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
@@ -377,9 +381,8 @@ if (GGML_CPU_ALL_VARIANTS)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-            ggml_add_cpu_backend_variant(s390x_z15  Z15 VXE)
-            # ggml_add_cpu_backend_variant(s390x_z16  Z16 VXE)
-            # ggml_add_cpu_backend_variant(s390x_z17  Z17 VXE)
+            ggml_add_cpu_backend_variant(z15    Z15 VXE2)
+            ggml_add_cpu_backend_variant(z16    Z16 VXE2 NNPA)
         else()
             message(FATAL_ERROR "Unsupported s390x target OS: ${CMAKE_SYSTEM_NAME}")
         endif()

ggml/src/ggml-cpu/CMakeLists.txt

@@ -504,11 +504,18 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             endforeach()
         endif()
 
-        if (GGML_VXE OR GGML_INTERNAL_VXE)
-            message(STATUS "VX/VXE/VXE2 enabled")
+        if (GGML_VXE OR GGML_INTERNAL_VXE2)
+            message(STATUS "VXE2 enabled")
             list(APPEND ARCH_FLAGS -mvx -mzvector)
-            list(APPEND ARCH_DEFINITIONS GGML_VXE)
+            list(APPEND ARCH_DEFINITIONS GGML_USE_VXE2)
         endif()
+
+        if (GGML_INTERNAL_NNPA)
+            message(STATUS "NNPA enabled")
+            list(APPEND ARCH_DEFINITIONS GGML_USE_NNPA)
+        endif()
+
+        ggml_add_cpu_backend_features(${GGML_CPU_NAME} s390 ${ARCH_DEFINITIONS})
     elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
         message(STATUS "Wasm detected")
         list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)

ggml/src/ggml-cpu/arch/loongarch/quants.c

@@ -700,7 +700,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     for (; ib + 1 < nb; ib += 2) {
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
+        const float ft0 = GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d);
+        const __m128 d_0_1 = (__m128)(v4f32){ft0, ft0, ft0, ft0};
 
         const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
 
@@ -714,11 +715,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         bx_1 = __lsx_vsub_b(bx_1, off);
         const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
 
-        //_mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
-        //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
-
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
+        const float ft1 = GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d);
+        const __m128 d_2_3 = (__m128)(v4f32){ft1, ft1, ft1, ft1};
 
         const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
 

ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp

@@ -0,0 +1,50 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__s390x__)
+#include <sys/auxv.h>
+
+// find hwcap bits in asm/elf.h
+#ifndef HWCAP_VXRS_EXT2
+#define HWCAP_VXRS_EXT2 (1 << 15)
+#endif
+
+#ifndef HWCAP_NNPA
+#define HWCAP_NNPA (1 << 20)
+#endif
+
+struct s390x_features {
+    bool has_vxe2 = false;
+    bool has_nnpa = false;
+
+    s390x_features() {
+        uint32_t hwcap = getauxval(AT_HWCAP);
+        // NOTE: use hwcap2 with DFLT for z17 and later
+        // uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+        has_vxe2 = !!(hwcap & HWCAP_VXRS_EXT2);
+        has_nnpa = !!(hwcap & HWCAP_NNPA);
+    }
+};
+
+static int ggml_backend_cpu_s390x_score() {
+    int score = 1;
+    s390x_features sf;
+
+// IBM z15 / LinuxONE 3
+#ifdef GGML_USE_VXE2
+    if (!sf.has_vxe2) { return 0; }
+    score += 1 << 1;
+#endif
+
+// IBM z16 / LinuxONE 4 and z17 / LinuxONE 5
+#ifdef GGML_USE_NNPA
+    if (!sf.has_nnpa) { return 0; }
+    score += 1 << 2;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_s390x_score)
+
+#endif  // __s390x__

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -500,13 +500,15 @@ inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
 
 #endif
 
-#if defined(__loongarch_asx)
+#if defined(__loongarch_sx)
 /* float type data load instructions */
 static __m128 __lsx_vreplfr2vr_s(const float val) {
     v4f32 res = {val, val, val, val};
     return (__m128)res;
 }
+#endif
 
+#if defined(__loongarch_asx)
 static __m256 __lasx_xvreplfr2vr_s(const float val) {
     v8f32 res = {val, val, val, val, val, val, val, val};
     return (__m256)res;

ggml/src/ggml-cpu/ops.cpp

@@ -7507,10 +7507,17 @@ static void ggml_compute_forward_upscale_f32(
     float sf1 = (float)ne1/src0->ne[1];
     float sf2 = (float)ne2/src0->ne[2];
     float sf3 = (float)ne3/src0->ne[3];
+    float pixel_offset = 0.5f;
 
     const int32_t mode_flags = ggml_get_op_params_i32(dst, 0);
     const ggml_scale_mode mode = (ggml_scale_mode) (mode_flags & 0xFF);
 
+    if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+        pixel_offset = 0.0f;
+        sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
+        sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
+    }
+
     if (mode == GGML_SCALE_MODE_NEAREST) {
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
@@ -7530,13 +7537,6 @@ static void ggml_compute_forward_upscale_f32(
             }
         }
     } else if (mode == GGML_SCALE_MODE_BILINEAR) {
-        float pixel_offset = 0.5f;
-        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
-            pixel_offset = 0.0f;
-            sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
-            sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
-        }
-
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
             for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
@@ -7571,6 +7571,51 @@ static void ggml_compute_forward_upscale_f32(
 
                         const float val = a*(1 - dx)*(1 - dy) + b*dx*(1 - dy) + c*(1 - dx)*dy + d*dx*dy;
 
+                        float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+                        *y_dst = val;
+                    }
+                }
+            }
+        }
+    } else if (mode == GGML_SCALE_MODE_BICUBIC) {
+        // https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
+        const float a = -0.75f; // use alpha = -0.75 (same as PyTorch)
+        auto weight1 = [a](float x) { return ((a + 2) * x - (a + 3)) * x * x + 1; };
+        auto weight2 = [a](float x) { return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a; };
+        auto bicubic = [=](float p0, float p1, float p2, float p3, float x) {
+            const float w0 = weight2(x + 1);
+            const float w1 = weight1(x + 0);
+            const float w2 = weight1(1 - x);
+            const float w3 = weight2(2 - x);
+            return p0*w0 + p1*w1 + p2*w2 + p3*w3;
+        };
+
+        for (int64_t i3 = 0; i3 < ne3; i3++) {
+            const int64_t i03 = i3 / sf3;
+            for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+                const int64_t i02 = i2 / sf2;
+                for (int64_t i1 = 0; i1 < ne1; i1++) {
+                    const float y = ((float)i1 + pixel_offset) / sf1 - pixel_offset;
+                    const int64_t y0 = (int64_t)floorf(y);
+                    const float dy = y - (float)y0;
+
+                    for (int64_t i0 = 0; i0 < ne0; i0++) {
+                        const float x = ((float)i0 + pixel_offset) / sf0 - pixel_offset;
+                        const int64_t x0 = (int64_t)floorf(x);
+                        const float dx = x - (float)x0;
+
+                        auto p = [=](int64_t x_off, int64_t y_off) -> float {
+                            int64_t i00 = std::max(int64_t(0), std::min(x0 + x_off, ne00 - 1));
+                            int64_t i01 = std::max(int64_t(0), std::min(y0 + y_off, ne01 - 1));
+                            return *(const float *)((const char *)src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                        };
+
+                        const float val = bicubic(
+                            bicubic(p(-1,-1), p(0,-1), p(1,-1), p(2,-1), dx),
+                            bicubic(p(-1, 0), p(0, 0), p(1, 0), p(2, 0), dx),
+                            bicubic(p(-1, 1), p(0, 1), p(1, 1), p(2, 1), dx),
+                            bicubic(p(-1, 2), p(0, 2), p(1, 2), p(2, 2), dx), dy);
+
                         float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
                         *y_dst = val;
                     }

ggml/src/ggml-cpu/repack.cpp

@@ -1678,10 +1678,24 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
         int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
 
+        // Ensure minimum chunk size to avoid alignment issues with high thread counts
+        // Minimum chunk size should be at least NB_COLS to prevent overlapping chunks after alignment
+        const int64_t min_chunk_size = NB_COLS;
+        if (nchunk > 0 && (nr / nchunk) < min_chunk_size && nr >= min_chunk_size) {
+            nchunk = (nr + min_chunk_size - 1) / min_chunk_size;
+        }
+
         if (nth == 1 || nchunk < nth || disable_chunking) {
             nchunk = nth;
         }
 
+        // Ensure nchunk doesn't exceed the number of rows divided by minimum chunk size
+        // This prevents creating too many tiny chunks that could overlap after alignment
+        const int64_t max_nchunk = (nr + min_chunk_size - 1) / min_chunk_size;
+        if (nchunk > max_nchunk) {
+            nchunk = max_nchunk;
+        }
+
         if (ith == 0) {
             // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
             ggml_threadpool_chunk_set(params->threadpool, nth);
@@ -1695,8 +1709,15 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         while (current_chunk < nchunk) {
             int64_t src0_start = (current_chunk * ne01) / nchunk;
             int64_t src0_end   = ((current_chunk + 1) * ne01) / nchunk;
+
+            // Align boundaries to NB_COLS - round up to ensure all data is included
+            // The chunk size limiting above ensures chunks are large enough to prevent overlaps
             src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
             src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
+            if (src0_end > ne01) {
+                src0_end = ne01;
+            }
+
             if (src0_start >= src0_end) {
                 break;
             }
@@ -1808,8 +1829,12 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             int64_t src0_cur_start = (ith * ne01) / nth;
             int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
 
+            // Align boundaries to NB_COLS - round up to ensure all data is included
             src0_cur_start = (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start;
             src0_cur_end   = (src0_cur_end   % NB_COLS) ? src0_cur_end   + NB_COLS - (src0_cur_end   % NB_COLS) : src0_cur_end;
+            if (src0_cur_end > ne01) {
+                src0_cur_end = ne01;
+            }
 
             if (src0_cur_start >= src0_cur_end) {
                 return;

ggml/src/ggml-cpu/simd-mappings.h

@@ -956,7 +956,7 @@ do {                                                              \
 
 #define GGML_F32Cx8          __m256
 #define GGML_F32Cx8_ZERO    (__m256)__lasx_xvldi(0)
-#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x))
+#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplfr2vr_s((x))
 
 static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) {
     __m256i a;
@@ -999,34 +999,34 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 
 #define GGML_F32x4         __m128
 #define GGML_F32x4_ZERO    (__m128)__lsx_vldi(0)
-#define GGML_F32x4_SET1(x) (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32x4_SET1(x) (__m128)__lsx_vreplfr2vr_s((x))
 #define GGML_F32x4_LOAD(x) (__m128)__lsx_vld((x), 0)
 #define GGML_F32x4_STORE(x, y)   __lsx_vst(y, x, 0)
 #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
 #define GGML_F32x4_ADD     __lsx_vfadd_s
 #define GGML_F32x4_MUL     __lsx_vfmul_s
-#define GGML_F32x4_REDUCE(res, x)                                                     \
-{                                                                                     \
-    int offset = GGML_F32_ARR >> 1;                                                   \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    offset >>= 1;                                                                     \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    offset >>= 1;                                                                     \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    __m128i tmp     = __lsx_vsrli_d((__m128i) x[0], 32);                              \
-    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]);                    \
-    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
-    const __m128 t0 = (__m128)__lsx_vshuf4i_w(tmp, 0x88);                                     \
-    tmp             = __lsx_vsrli_d((__m128i) t0, 32);                                \
-    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, t0);                      \
-    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
-    res             = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \
+
+#define GGML_F32x4_REDUCE(res, x)                               \
+{                                                               \
+    int offset = GGML_F32_ARR >> 1;                             \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    offset >>= 1;                                               \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    offset >>= 1;                                               \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    __m128i t0 = __lsx_vpickev_w((__m128i)x[0], (__m128i)x[0]); \
+    __m128i t1 = __lsx_vpickod_w((__m128i)x[0], (__m128i)x[0]); \
+    __m128 t2 = __lsx_vfadd_s((__m128)t0, (__m128)t1);          \
+    __m128i t3 = __lsx_vpickev_w((__m128i)t2, (__m128i)t2);     \
+    __m128i t4 = __lsx_vpickod_w((__m128i)t2, (__m128i)t2);     \
+    __m128 t5 = __lsx_vfadd_s((__m128)t3, (__m128)t4);          \
+    res = (ggml_float) ((v4f32)t5)[0];                          \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -1068,7 +1068,7 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 
 #define GGML_F32Cx4             __m128
 #define GGML_F32Cx4_ZERO        (__m128)__lsx_vldi(0)
-#define GGML_F32Cx4_SET1(x)     (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32Cx4_SET1(x)     (__m128)__lsx_vreplfr2vr_s((x))
 #define GGML_F32Cx4_LOAD(x)     (__m128)__lsx_f16x4_load(x)
 #define GGML_F32Cx4_STORE(x, y) __lsx_f16x4_store(x, y)
 #define GGML_F32Cx4_FMA         GGML_F32x4_FMA

ggml/src/ggml-cuda/common.cuh

@@ -224,6 +224,11 @@ static const char * cu_get_error_str(CUresult err) {
 #define AMD_MFMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
+// The Volta instructions are in principle available on Turing or newer but they are effectively unusable:
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#define VOLTA_MMA_AVAILABLE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define TURING_MMA_AVAILABLE
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
@@ -278,7 +283,10 @@ static bool amd_mfma_available(const int cc) {
 #endif //!defined(GGML_HIP_NO_MMQ_MFMA)
 }
 
-// Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
+static bool volta_mma_available(const int cc) {
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_VOLTA;
+}
+
 static bool turing_mma_available(const int cc) {
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
 }

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

@@ -14,6 +14,10 @@ void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_cuda_flash_attn_ext_tile_case< 64,  64>(ctx, dst);
         } break;
+        case  72: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 72,  72>(ctx, dst);
+        } break;
         case  80: {
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_cuda_flash_attn_ext_tile_case< 80,  80>(ctx, dst);

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

@@ -6,7 +6,7 @@
 // nbatch_K == number of K columns to load in parallel for KQ calculation
 
 // TODO optimize kernel parameters for FP16 NVIDIA (P100)
-// TODO optimize kernel parameters for head sizes 40, 80, 96, 112
+// TODO optimize kernel parameters for head sizes 40, 72, 80, 96, 112
 
 // The ROCm compiler cannot handle templating in __launch_bounds__.
 // As a workaround, define a macro to package the kernel parameters as uint32_t:
@@ -32,6 +32,12 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  64,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  64,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  64,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  64,  40)
@@ -80,6 +86,12 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 128, 3,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
@@ -130,6 +142,13 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
@@ -185,6 +204,13 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 128, 4,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 128, 5,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
@@ -723,7 +749,7 @@ static __global__ void flash_attn_tile(
 
     if (
 #ifdef GGML_USE_WMMA_FATTN
-            (ncols2 != 1 && DV != 40 && DV != 512) ||
+            (ncols2 != 1 && DV != 40 && DV != 72 && DV != 512) ||
 #endif // GGML_USE_WMMA_FATTN
             (use_logit_softcap && !(DV == 128 || DV == 256))
     ) {
@@ -1198,6 +1224,7 @@ void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor
 
 extern DECL_FATTN_TILE_CASE( 40,  40);
 extern DECL_FATTN_TILE_CASE( 64,  64);
+extern DECL_FATTN_TILE_CASE( 72,  72);
 extern DECL_FATTN_TILE_CASE( 80,  80);
 extern DECL_FATTN_TILE_CASE( 96,  96);
 extern DECL_FATTN_TILE_CASE(112, 112);

ggml/src/ggml-cuda/fattn.cu

@@ -223,6 +223,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     switch (K->ne[0]) {
         case  40:
         case  64:
+        case  72:
         case  80:
         case  96:
         case 128:
@@ -275,7 +276,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
     // If Turing tensor cores available, use them:
-    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40) {
+    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72) {
         if (can_use_vector_kernel) {
             if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE && Q->ne[1] == 1 && Q->ne[3] == 1 && !(gqa_ratio > 4 && K->ne[1] >= 8192)) {
@@ -301,7 +302,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     }
 
     // Use the WMMA kernel if possible:
-    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 576) {
+    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 576) {
         if (can_use_vector_kernel && Q->ne[1] <= 2) {
             return BEST_FATTN_KERNEL_VEC;
         }

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

@@ -27,6 +27,7 @@
 #include "ggml-cuda/mmq.cuh"
 #include "ggml-cuda/mmvf.cuh"
 #include "ggml-cuda/mmvq.cuh"
+#include "ggml-cuda/moe-expert-reduce.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
 #include "ggml-cuda/opt-step-sgd.cuh"
@@ -2114,6 +2115,14 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor) {
     const int cc      = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
 
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
     //we only support fusion for ncols_dst = 1
     if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
         return false;
@@ -2153,6 +2162,15 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
         return false;
     }
 
+
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
     return use_mul_mat_vec_q;
 }
 
@@ -2498,6 +2516,18 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_XIELU:
                     ggml_cuda_op_xielu(ctx, dst);
                     break;
+                case GGML_UNARY_OP_FLOOR:
+                    ggml_cuda_op_floor(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_CEIL:
+                    ggml_cuda_op_ceil(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_ROUND:
+                    ggml_cuda_op_round(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_TRUNC:
+                    ggml_cuda_op_trunc(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -3169,6 +3199,31 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                         continue;
                     }
 
+                    if (node->op == GGML_OP_MUL) {
+                        int current_node = i + 1;
+                        int num_views    = 0;
+                        int num_adds     = 0;
+                        while (current_node < cgraph->n_nodes && cgraph->nodes[current_node]->op == GGML_OP_VIEW) {
+                            num_views++;
+                            current_node++;
+                        }
+
+                        while (current_node < cgraph->n_nodes && cgraph->nodes[current_node]->op == GGML_OP_ADD &&
+                                num_adds < num_views - 1) {
+                            num_adds++;
+                            current_node++;
+                        }
+
+                        if (num_adds == num_views - 1 && num_views > 0) {
+                            ggml_tensor * dst_node = cgraph->nodes[current_node - 1];
+                            if (ggml_cuda_should_use_moe_expert_reduce(cgraph, i, current_node)) {
+                                ggml_cuda_op_moe_expert_reduce(*cuda_ctx, node->src[0], node->src[1], dst_node);
+                                i += num_views + num_adds;
+                                continue;
+                            }
+                        }
+                    }
+
                     if (node->op == GGML_OP_ADD) {
                         int n_fuse = 0;
                         ggml_op ops[8];
@@ -3743,6 +3798,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
                 case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;

ggml/src/ggml-cuda/mma.cuh

@@ -18,6 +18,10 @@
 
 #include "common.cuh"
 
+// On Volta each warp is doing 4 8x8 mma operations in parallel.
+// The basic memory layout for a 32x8 output tile is to stack 4 input tiles in I direction and to mirror the B tile.
+// However, the i indices in this file are by default permuted to simplify the index calculations.
+// #define GGML_CUDA_MMA_NO_VOLTA_PERM
 
 #if CUDART_VERSION >= 11080
 
@@ -73,6 +77,15 @@ namespace ggml_cuda_mma {
         static constexpr int ne = I * J / 64;
         T x[ne] = {0};
 
+        static constexpr __device__ bool supported() {
+            if (I == 64 && J ==  2) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 32 && J ==  4) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J == 32) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
                 return threadIdx.x % 16;
@@ -85,7 +98,8 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 32 && J == 32) {
                 return 4 * (threadIdx.x / 32) + 8 * (l / 4) + (l % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
@@ -101,22 +115,67 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 32 && J == 32) {
                 return threadIdx.x % 32;
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        static constexpr int ne = I * J / 32;
+        T x[ne] = {0};
+
+        static constexpr __device__ bool supported() {
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 32 && J == 8) {
+#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
+                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (l & 2) | (threadIdx.x % 2);
+#else
+                return (l & 2) | (threadIdx.x & ~2);
+#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 32 && J == 8) {
+                return (threadIdx.x & 2) | (l & (4 + 1));
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 #else
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 8 && (J == 4 || J == 8)) {
+            if constexpr (I == 8 && J == 4) {
+                return threadIdx.x / 4;
+            } else if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 8 + threadIdx.x / 4;
+                return ((l / 2) * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 16) {
-                return ((l / 2) % 2) * 8 + threadIdx.x / 4;
+                return (((l / 2) % 2) * 8) | (threadIdx.x / 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return tile<16, 8, T>::get_i(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
@@ -124,13 +183,16 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 8 && J == 8) {
-                return 4 * l + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 8) {
-                return 2 * (threadIdx.x % 4) + l % 2;
+                return ((threadIdx.x % 4) * 2) | (l % 2);
             } else if constexpr (I == 16 && J == 16) {
-                return 8 * (l / 4) + 2 * (threadIdx.x % 4) + l % 2;
+                return ((l / 4) * 8) | ((threadIdx.x % 4) * 2) | (l % 2);
+            } else if constexpr (I == 32 && J == 8) {
+                return tile<16, 8, T>::get_j(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 #endif // defined(GGML_USE_HIP)
@@ -140,32 +202,83 @@ namespace ggml_cuda_mma {
     struct tile<I_, J_, half2> {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
+
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        static constexpr int ne = I == 8 && J == 8 ? I * J / (WARP_SIZE/4) : I * J / WARP_SIZE;
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  8) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 8 && J == 8) {
+                return ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+            } else if constexpr (I == 32 && J == 8) {
+#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
+                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+#else
+                return threadIdx.x;
+#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr ((I == 8 || I == 32) && J == 8) {
+                return l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#else
         static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return l * 8 + threadIdx.x / 4;
+                return (l * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return (l % 2) * 8 + threadIdx.x / 4;
+                return ((l % 2) * 8) | (threadIdx.x / 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return ((l / 4) * 16) | ((l % 2) * 8) | (threadIdx.x / 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return l * 4 + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 4 + threadIdx.x % 4;
+                return ((l / 2) * 4) | (threadIdx.x % 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return ((l & 2) * 2) | (threadIdx.x % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
     };
 
     template <int I_, int J_>
@@ -175,27 +288,36 @@ namespace ggml_cuda_mma {
         static constexpr int ne = I * J / WARP_SIZE;
         nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  4) return true;
+            if (I == 16 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return l * 8 + threadIdx.x / 4;
+                return (l * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return (l % 2) * 8 + threadIdx.x / 4;
+                return ((l % 2) * 8) | (threadIdx.x / 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return l * 4 + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 4 + threadIdx.x % 4;
+                return ((l / 2) * 4) | (threadIdx.x % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
     };
@@ -263,8 +385,12 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
 #else
-        load_generic(xs0, stride);
-        GGML_UNUSED(t);
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
+#else
+        load_generic(t, xs0, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -277,11 +403,35 @@ namespace ggml_cuda_mma {
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
             : "l"(xs));
+#else
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #else
         load_generic(t, xs0, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
+    template <typename T>
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<32, 8, T> & t, const T * __restrict__ xs0, const int stride) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if 1
+        // TODO: more generic handling
+        static_assert(sizeof(T) == 4, "bad type size");
+        ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 4, xs0 + t.get_i(4)*stride + 4);
+#else
+        load_generic(t, xs0, stride);
+#endif // 1
+#else
+        tile<16, 8, T> * t16 = (tile<16, 8, T> *) &t;
+        load_ldmatrix(t16[0], xs0 +  0*stride, stride);
+        load_ldmatrix(t16[1], xs0 + 16*stride, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+    }
+
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix_trans(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
@@ -546,4 +696,43 @@ namespace ggml_cuda_mma {
         NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
     }
+
+    template <typename T1, typename T2, int J, int K>
+    static __device__ __forceinline__ void mma(
+            tile<32, J, T1> & D, const tile<32, K, T2> & A, const tile<J, K, T2> & B) {
+        tile<16, J, T1> * D16 = (tile<16, J, T1> *) &D;
+        tile<16, K, T2> * A16 = (tile<16, K, T2> *) &A;
+        mma(D16[0], A16[0], B);
+        mma(D16[1], A16[1], B);
+    }
+
+    static __device__ __forceinline__ void mma(
+            tile<32, 8, float> & D, const tile<32, 8, half2> & A, const tile<8, 8, half2> & B) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0]), "r"(Bxi[1]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]), "r"(Bxi[3]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[4]), "r"(Axi[5]), "r"(Bxi[4]), "r"(Bxi[5]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7]));
+#else
+        tile<16, 8, float> * D16 = (tile<16, 8, float> *) &D;
+        tile<16, 8, half2> * A16 = (tile<16, 8, half2> *) &A;
+        mma(D16[0], A16[0], B);
+        mma(D16[1], A16[1], B);
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+    }
 }

ggml/src/ggml-cuda/mmf.cu

@@ -148,7 +148,7 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
         case GGML_TYPE_F32:
             return ampere_mma_available(cc);
         case GGML_TYPE_F16:
-            return turing_mma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc);
         case GGML_TYPE_BF16:
             return ampere_mma_available(cc);
         default:

ggml/src/ggml-cuda/mmf.cuh

@@ -28,9 +28,19 @@ static __global__ void mul_mat_f(
         const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
+    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
+    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
+
+    if (!I_16_supported && !I_32_supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work but 16 is ~1% faster.
+
+    typedef tile<I_preferred, 8, T>     tile_A;
+    typedef tile<8,           8, T>     tile_B;
+    typedef tile<I_preferred, 8, float> tile_C;
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -232,7 +242,6 @@ static __global__ void mul_mat_f(
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
-
 //This kernel is for larger batch sizes of mul_mat_id
 template <typename T, int rows_per_block, int cols_per_block, int nwarps>
 __launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
@@ -245,9 +254,19 @@ static __global__ void mul_mat_f_ids(
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
         const uint3 sis1_fd, const uint3 nch_fd) {
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
+    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
+    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
+
+    if (!I_16_supported && !I_32_supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work butr 16 is ~1% faster.
+
+    typedef tile<I_preferred, 8, T>     tile_A;
+    typedef tile<8,           8, T>     tile_B;
+    typedef tile<I_preferred, 8, float> tile_C;
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -533,7 +552,8 @@ void mul_mat_f_cuda(
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         cudaStream_t stream, const mmf_ids_data * ids_data) {
-    typedef tile<16, 8, T>     tile_A;
+    typedef tile<16, 8, T>     tile_A_16;
+    typedef tile<32, 8, T>     tile_A_32;
     typedef tile< 8, 8, T>     tile_B;
 
     GGML_ASSERT(ncols_x      % 2 == 0);
@@ -544,7 +564,8 @@ void mul_mat_f_cuda(
     const int64_t channel_ratio = nchannels_dst / nchannels_x;
     const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
 
-    const int device = ggml_cuda_get_device();
+    const int device    = ggml_cuda_get_device();
+    const int cc        = ggml_cuda_info().devices[device].cc;
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
 
     int64_t nwarps_best     = 1;
@@ -559,7 +580,7 @@ void mul_mat_f_cuda(
     }
 
     constexpr int rows_per_block = MMF_ROWS_PER_BLOCK;
-    const int nbytes_shared_iter = nwarps_best * tile_A::I * (warp_size + 4) * 4;
+    const int nbytes_shared_iter = nwarps_best * (volta_mma_available(cc) ? tile_A_32::I : tile_A_16::I) * (warp_size + 4) * 4;
     const int nbytes_shared_combine = GGML_PAD(cols_per_block, tile_B::I) * (nwarps_best*rows_per_block + 4) * 4;
     const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
     const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0;

ggml/src/ggml-cuda/mmvq.cu

@@ -190,8 +190,8 @@ static __global__ void mul_mat_vec_q(
 
     const uint32_t channel_bias = ids ? channel_x : channel_dst;
 
-    float x_biases[ncols_dst][rows_per_cuda_block]    = { { 0.0f } };
-    float gate_biases[ncols_dst][rows_per_cuda_block] = { { 0.0f } };
+    float x_biases[ncols_dst]    = { 0.0f };
+    float gate_biases[ncols_dst] = { 0.0f };
     if constexpr (has_fusion) {
         if (use_bias) {
             x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
@@ -199,8 +199,9 @@ static __global__ void mul_mat_vec_q(
             // 2. load only on threads that won't die after partial sum calculation
             if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
                 (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+#pragma unroll
                 for (int j = 0; j < ncols_dst; ++j) {
-                    x_biases[j][threadIdx.x] = x_bias[j * stride_col_dst + threadIdx.x];
+                    x_biases[j] = x_bias[j * stride_col_dst + threadIdx.x];
                 }
             }
         }
@@ -208,8 +209,9 @@ static __global__ void mul_mat_vec_q(
             gate_bias = gate_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
             if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
                 (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+#pragma unroll
                 for (int j = 0; j < ncols_dst; ++j) {
-                    gate_biases[j][threadIdx.x] = gate_bias[j * stride_col_dst + threadIdx.x];
+                    gate_biases[j] = gate_bias[j * stride_col_dst + threadIdx.x];
                 }
             }
         }
@@ -299,12 +301,12 @@ static __global__ void mul_mat_vec_q(
             float result = tmp[j][threadIdx.x];
             if constexpr (has_fusion) {
                 if (use_bias) {
-                    result += x_biases[j][threadIdx.x];
+                    result += x_biases[j];
                 }
                 if (use_gate) {
                     float gate_value = tmp_gate[j][threadIdx.x];
                     if (use_gate_bias) {
-                        gate_value += gate_biases[j][threadIdx.x];
+                        gate_value += gate_biases[j];
                     }
                     switch (active_glu) {
                         case GGML_GLU_OP_SWIGLU:

ggml/src/ggml-cuda/moe-expert-reduce.cu

@@ -0,0 +1,168 @@
+#include "moe-expert-reduce.cuh"
+
+// This kernel is a fusion of the expert weight reduce, common in MoE models
+
+template <int n_expert_used_template>
+__global__ void moe_expert_reduce_cuda(const float * __restrict__ experts,
+                                       const float * __restrict__ weights,
+                                       float * __restrict__ dst,
+                                       const int n_expert_used,
+                                       const int n_cols) {
+    const int row = blockIdx.x;
+    const int col = blockIdx.y * blockDim.x + threadIdx.x;
+    if (col >= n_cols) {
+        return;
+    }
+
+    experts += row * n_cols * n_expert_used;
+    weights += row * n_expert_used;
+    dst += row * n_cols;
+
+    float acc = 0.f;
+    if constexpr (n_expert_used_template == 0) {
+        for (int expert = 0; expert < n_expert_used; ++expert) {
+            ggml_cuda_mad(acc, experts[col], weights[expert]);
+            experts += n_cols;
+        }
+        dst[col] = acc;
+    } else {
+#pragma unroll
+        for (int i = 0; i < n_expert_used_template; ++i) {
+            ggml_cuda_mad(acc, experts[col], weights[i]);
+            experts += n_cols;
+        }
+        dst[col] = acc;
+    }
+}
+
+static void launch_moe_expert_reduce(ggml_backend_cuda_context & ctx,
+                                     const float *               experts,
+                                     const float *               weights,
+                                     float *                     dst,
+                                     const int                   n_expert_used,
+                                     const int                   n_cols,
+                                     const int                   n_rows) {
+    const int block_size = 32;
+
+    const int n_blocks_x = n_rows;
+    const int n_blocks_y = (n_cols + block_size - 1) / block_size;
+
+    dim3 block_dims(block_size);
+    dim3 grid_dims(n_blocks_x, n_blocks_y);
+
+    cudaStream_t stream = ctx.stream();
+    switch (n_expert_used) {
+        case 1:
+            moe_expert_reduce_cuda<1>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 2:
+            moe_expert_reduce_cuda<2>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 4:
+            moe_expert_reduce_cuda<4>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 6:
+            moe_expert_reduce_cuda<6>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 8:
+            moe_expert_reduce_cuda<8>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 16:
+            moe_expert_reduce_cuda<16>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 32:
+            moe_expert_reduce_cuda<32>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 64:
+            moe_expert_reduce_cuda<64>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 128:
+            moe_expert_reduce_cuda<128>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        default:
+            moe_expert_reduce_cuda<0>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+    }
+}
+
+bool ggml_cuda_should_use_moe_expert_reduce(const ggml_cgraph * cgraph, int start_index, int end_index) {
+    const ggml_tensor * mul = cgraph->nodes[start_index];
+
+    if (mul->op != GGML_OP_MUL || !ggml_is_contiguous(mul->src[0]) || !ggml_is_contiguous(mul->src[1])) {
+        return false;
+    }
+
+    int    current_node   = start_index + 1;
+    size_t current_offset = 0;
+
+    std::vector<const ggml_tensor *> view_nodes;
+    //check if all are views of the expert in increasing order
+    while (current_node < end_index && cgraph->nodes[current_node]->op == GGML_OP_VIEW) {
+        const ggml_tensor * node = cgraph->nodes[current_node];
+        if (node->view_src != mul) {
+            return false;
+        }
+        if (node->view_offs < current_offset) {
+            return false;
+        }
+        current_offset = node->view_offs;
+        current_node++;
+        view_nodes.push_back(node);
+    }
+
+    //check if all the adds are in increasing order
+    const ggml_tensor * prev_add_src = view_nodes.empty() ? nullptr : view_nodes[0];
+    int                 num_adds     = 0;
+    int                 num_views    = view_nodes.size();
+    while (current_node < end_index && cgraph->nodes[current_node]->op == GGML_OP_ADD) {
+        const ggml_tensor * add_node = cgraph->nodes[current_node];
+
+        bool is_first_op_ok  = num_views > num_adds ? add_node->src[0] == prev_add_src : false;
+        bool is_second_op_ok = num_views > num_adds ? add_node->src[1] == view_nodes[num_adds + 1] : false;
+
+        if (!is_first_op_ok || !is_second_op_ok) {
+            return false;
+        }
+        prev_add_src = add_node;
+
+        num_adds++;
+        current_node++;
+    }
+
+    if (num_views != num_adds + 1) {
+        return false;
+    }
+
+    return true;
+}
+
+void ggml_cuda_op_moe_expert_reduce(ggml_backend_cuda_context & ctx,
+                                    const ggml_tensor *         experts,
+                                    const ggml_tensor *         weights,
+                                    ggml_tensor *               dst) {
+    const int n_rows        = experts->ne[2];
+    const int n_expert_used = experts->ne[1];
+    const int n_cols        = experts->ne[0];
+
+    GGML_ASSERT(experts->type == GGML_TYPE_F32);
+    GGML_ASSERT(weights->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(experts));
+    GGML_ASSERT(ggml_is_contiguous(weights));
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const float * experts_d = (const float *) experts->data;
+    const float * weights_d = (const float *) weights->data;
+    float *       dst_d     = (float *) dst->data;
+
+    launch_moe_expert_reduce(ctx, experts_d, weights_d, dst_d, n_expert_used, n_cols, n_rows);
+}

ggml/src/ggml-cuda/moe-expert-reduce.cuh

@@ -0,0 +1,11 @@
+#include "common.cuh"
+#include "ggml.h"
+
+#include <initializer_list>
+
+void ggml_cuda_op_moe_expert_reduce(ggml_backend_cuda_context & ctx,
+                                    const ggml_tensor *         experts,
+                                    const ggml_tensor *         weights,
+                                    ggml_tensor *               dst);
+
+bool ggml_cuda_should_use_moe_expert_reduce(const ggml_cgraph * cgraph, int start_index, int end_index);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(72, 72);

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

@@ -3,7 +3,7 @@
 from glob import glob
 import os
 
-HEAD_SIZES_KQ = [40, 64, 80, 96, 112, 128, 256, 576]
+HEAD_SIZES_KQ = [40, 64, 72, 80, 96, 112, 128, 256, 576]
 
 TYPES_KV = ["GGML_TYPE_F16", "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0"]
 
@@ -81,6 +81,8 @@ for ncols in [8, 16, 32, 64]:
             for head_size_kq in HEAD_SIZES_KQ:
                 if head_size_kq == 40:
                     continue
+                if head_size_kq == 72:
+                    continue
                 if head_size_kq != 576 and ncols2 == 16:
                     continue
                 if head_size_kq == 576 and ncols2 != 16:

ggml/src/ggml-cuda/unary.cu

@@ -85,6 +85,22 @@ static __device__ __forceinline__ float op_elu(float x) {
     return (x > 0.f) ? x : expm1f(x);
 }
 
+static __device__ __forceinline__ float op_floor(float x) {
+    return floorf(x);
+}
+
+static __device__ __forceinline__ float op_ceil(float x) {
+    return ceilf(x);
+}
+
+static __device__ __forceinline__ float op_round(float x) {
+    return round(x);
+}
+
+static __device__ __forceinline__ float op_trunc(float x) {
+    return trunc(x);
+}
+
 template <float (*op)(float), typename T>
 static __global__ void unary_op_kernel(const T * x, T * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
@@ -201,6 +217,22 @@ void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_unary<op_elu>(ctx, dst);
 }
+
+void ggml_cuda_op_floor(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_floor>(ctx, dst);
+}
+
+void ggml_cuda_op_ceil(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_ceil>(ctx, dst);
+}
+
+void ggml_cuda_op_round(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_round>(ctx, dst);
+}
+
+void ggml_cuda_op_trunc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_trunc>(ctx, dst);
+}
 /* gated ops */
 
 template <float (*op)(float), typename T>

ggml/src/ggml-cuda/unary.cuh

@@ -63,6 +63,14 @@ void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_floor(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_ceil(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_round(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_trunc(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -707,6 +707,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             if (op->src[0]->ne[0] != 32 &&
                 op->src[0]->ne[0] != 40 &&
                 op->src[0]->ne[0] != 64 &&
+                op->src[0]->ne[0] != 72 &&
                 op->src[0]->ne[0] != 80 &&
                 op->src[0]->ne[0] != 96 &&
                 op->src[0]->ne[0] != 112 &&

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

@@ -5362,6 +5362,7 @@ typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, hal
 template [[host_name("kernel_flash_attn_ext_f32_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  32,  32>;
 template [[host_name("kernel_flash_attn_ext_f32_dk40_dv40"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  40,  40>;
 template [[host_name("kernel_flash_attn_ext_f32_dk64_dv64"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  64,  64>;
+template [[host_name("kernel_flash_attn_ext_f32_dk72_dv72"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  72,  72>;
 template [[host_name("kernel_flash_attn_ext_f32_dk80_dv80"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  80,  80>;
 template [[host_name("kernel_flash_attn_ext_f32_dk96_dv96"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  96,  96>;
 template [[host_name("kernel_flash_attn_ext_f32_dk112_dv112")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  112, 112>;
@@ -5374,6 +5375,7 @@ template [[host_name("kernel_flash_attn_ext_f32_dk576_dv512")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  32,  32>;
 template [[host_name("kernel_flash_attn_ext_f16_dk40_dv40"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  40,  40>;
 template [[host_name("kernel_flash_attn_ext_f16_dk64_dv64"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  64,  64>;
+template [[host_name("kernel_flash_attn_ext_f16_dk72_dv72"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  72,  72>;
 template [[host_name("kernel_flash_attn_ext_f16_dk80_dv80"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  80,  80>;
 template [[host_name("kernel_flash_attn_ext_f16_dk96_dv96"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  96,  96>;
 template [[host_name("kernel_flash_attn_ext_f16_dk112_dv112")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  112, 112>;
@@ -5387,6 +5389,7 @@ template [[host_name("kernel_flash_attn_ext_f16_dk576_dv512")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_bf16_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_bf16_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
@@ -5400,6 +5403,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q4_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 112, 112>;
@@ -5412,6 +5416,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q4_1_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 112, 112>;
@@ -5424,6 +5429,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q5_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 112, 112>;
@@ -5436,6 +5442,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q5_1_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 112, 112>;
@@ -5448,6 +5455,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q8_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 112, 112>;

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

@@ -8399,6 +8399,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     const bool is_neox = mode & 2;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+    const int  is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
 
     if (is_mrope) {
         GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
@@ -8489,9 +8490,14 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     CL_CHECK(clSetKernelArg(kernel, 30, sizeof(float),    &attn_factor));
     CL_CHECK(clSetKernelArg(kernel, 31, sizeof(float),    &beta_fast));
     CL_CHECK(clSetKernelArg(kernel, 32, sizeof(float),    &beta_slow));
+    // both mrope and vision kernels have sections
     if (is_mrope || is_vision) {
         CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int32_t)*4, &sections));
     }
+    // only mrope has is_imrope
+    if (is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 34, sizeof(int), &is_imrope));
+    }
 
     size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
     size_t local_work_size[] = {(size_t)nth, 1, 1};

ggml/src/ggml-opencl/kernels/rope.cl

@@ -392,7 +392,8 @@ kernel void kernel_rope_multi_f32(
         float attn_factor,
         float beta_fast,
         float beta_slow,
-        int4 sections
+        int4 sections,
+        int  is_imrope
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -419,17 +420,29 @@ kernel void kernel_rope_multi_f32(
             const int sector = (i0 / 2) % sect_dims;
             float theta_base = 0.0f;
 
-            if (sector < sections.s0) {
-                theta_base = pos[i2];
-            }
-            else if (sector >= sections.s0 && sector < sec_w) {
-                theta_base = pos[i2 + ne2 * 1];
-            }
-            else if (sector >= sec_w && sector < sec_w + sections.s2) {
-                theta_base = pos[i2 + ne2 * 2];
-            }
-            else if (sector >= sec_w + sections.s2) {
-                theta_base = pos[i2 + ne2 * 3];
+            if (is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * sections.s1) { // h
+                    theta_base = (float) pos[i2 + ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * sections.s2) { // w
+                    theta_base = (float) pos[i2 + ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * sections.s0) { // t
+                    theta_base = (float) pos[i2 + ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + ne02 * 3];
+                }
+            } else {
+                if (sector < sections.s0) {
+                    theta_base = pos[i2];
+                }
+                else if (sector >= sections.s0 && sector < sec_w) {
+                    theta_base = pos[i2 + ne2 * 1];
+                }
+                else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 2];
+                }
+                else if (sector >= sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 3];
+                }
             }
 
             const float theta = theta_base * pow(freq_base, inv_ndims*i0);
@@ -490,7 +503,8 @@ kernel void kernel_rope_multi_f16(
         float attn_factor,
         float beta_fast,
         float beta_slow,
-        int4 sections
+        int4 sections,
+        int  is_imrope
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -517,17 +531,29 @@ kernel void kernel_rope_multi_f16(
             const int sector = (i0 / 2) % sect_dims;
             float theta_base = 0.0f;
 
-            if (sector < sections.s0) {
-                theta_base = pos[i2];
-            }
-            else if (sector >= sections.s0 && sector < sec_w) {
-                theta_base = pos[i2 + ne2 * 1];
-            }
-            else if (sector >= sec_w && sector < sec_w + sections.s2) {
-                theta_base = pos[i2 + ne2 * 2];
-            }
-            else if (sector >= sec_w + sections.s2) {
-                theta_base = pos[i2 + ne2 * 3];
+            if (is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * sections.s1) { // h
+                    theta_base = (float) pos[i2 + ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * sections.s2) { // w
+                    theta_base = (float) pos[i2 + ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * sections.s0) { // t
+                    theta_base = (float) pos[i2 + ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + ne02 * 3];
+                }
+            } else {
+                if (sector < sections.s0) {
+                    theta_base = pos[i2];
+                }
+                else if (sector >= sections.s0 && sector < sec_w) {
+                    theta_base = pos[i2 + ne2 * 1];
+                }
+                else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 2];
+                }
+                else if (sector >= sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 3];
+                }
             }
 
             const float theta = theta_base * pow(freq_base, inv_ndims*i0);

ggml/src/ggml-sycl/repeat_back.cpp

@@ -2,26 +2,43 @@
 
 #include "common.hpp"
 
-void ggml_sycl_op_repeat_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+#define GGML_ASSERT_TENSOR_FITS_INT(t) \
+    GGML_ASSERT((t)->ne[0] < INT_MAX && (t)->ne[1] < INT_MAX && (t)->ne[2] < INT_MAX && (t)->ne[3] < INT_MAX)
 
+void ggml_sycl_op_repeat_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     const float * src0_dd = (const float *) dst->src[0]->data;
     float *       dst_dd  = (float *) dst->data;
 
-    const int64_t ne0 = dst->ne[0], ne1 = dst->ne[1], ne2 = dst->ne[2], ne3 = dst->ne[3];
-    const int64_t ne00 = dst->src[0]->ne[0], ne01 = dst->src[0]->ne[1], ne02 = dst->src[0]->ne[2],
-                  ne03 = dst->src[0]->ne[3];
+    GGML_ASSERT_TENSOR_FITS_INT(dst);
+    GGML_ASSERT_TENSOR_FITS_INT(dst->src[0]);
 
-    const int nr0 = (int) (ne00 / ne0);
-    const int nr1 = (int) (ne01 / ne1);
-    const int nr2 = (int) (ne02 / ne2);
-    const int nr3 = (int) (ne03 / ne3);
+    const int ne0 = dst->ne[0], ne1 = dst->ne[1], ne2 = dst->ne[2], ne3 = dst->ne[3];
+    const int ne00 = dst->src[0]->ne[0], ne01 = dst->src[0]->ne[1], ne02 = dst->src[0]->ne[2],
+              ne03 = dst->src[0]->ne[3];
 
-    const size_t total      = ne0 * ne1 * ne2 * ne3;
-    const int    BLOCK_SIZE = 256;
-    const int    num_blocks = (total + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    const int nr0 = ne00 / ne0;
+    const int nr1 = ne01 / ne1;
+    const int nr2 = ne02 / ne2;
+    const int nr3 = ne03 / ne3;
+
+    const int nb0 = dst->src[0]->nb[0];
+    const int nb1 = dst->src[0]->nb[1];
+    const int nb2 = dst->src[0]->nb[2];
+    const int nb3 = dst->src[0]->nb[3];
+
+    const char * base = (const char *) src0_dd;
+
+    const size_t  total      = (size_t) ne0 * ne1 * ne2 * ne3;
+    constexpr int BLOCK_SIZE = 256;
+    const int     num_blocks = (total + BLOCK_SIZE - 1) / BLOCK_SIZE;
+
+    const float inv_ne0      = 1.0f / ne0;
+    const float inv_ne_01    = 1.0f / (ne0 * ne1);
+    const float inv_ne_012   = 1.0f / (ne0 * ne1 * ne2);
+    const int   repeat_count = nr0 * nr1 * nr2 * nr3;
 
     queue_ptr stream = ctx.stream();
 
@@ -33,24 +50,27 @@ void ggml_sycl_op_repeat_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst
                 return;
             }
 
-            const int i0 = i % ne0;
-            const int i1 = (i / ne0) % ne1;
-            const int i2 = (i / (ne0 * ne1)) % ne2;
-            const int i3 = i / (ne0 * ne1 * ne2);
+            const int i3 = (int) (i * inv_ne_012);
+            const int i2 = (int) (i * inv_ne_01) - i3 * ne2;
+            const int i1 = (int) (i * inv_ne0) - (int) (i * inv_ne_01) * ne1;
+            const int i0 = i - (int) (i * inv_ne0) * ne0;
 
+            int   j0 = 0, j1 = 0, j2 = 0, j3 = 0;
             float acc = 0.0f;
 
-            for (int j3 = 0; j3 < nr3; ++j3) {
-                for (int j2 = 0; j2 < nr2; ++j2) {
-                    for (int j1 = 0; j1 < nr1; ++j1) {
-                        for (int j0 = 0; j0 < nr0; ++j0) {
-                            acc += src0_dd[(i0 + j0 * ne0) + (i1 + j1 * ne1) * ne00 + (i2 + j2 * ne2) * ne00 * ne01 +
-                                           (i3 + j3 * ne3) * ne00 * ne01 * ne02];
-                        }
-                    }
-                }
-            }
+            for (int j = 0; j < repeat_count; ++j) {
+                const float * ptr = (const float *) (base + (i0 + j0 * ne0) * nb0 + (i1 + j1 * ne1) * nb1 +
+                    (i2 + j2 * ne2) * nb2 + (i3 + j3 * ne3) * nb3);
+                acc += *ptr;
 
+                int carry = (++j0 >= nr0);
+                j0 -= carry * nr0;
+                carry = (carry && (++j1 >= nr1));
+                j1 -= carry * nr1;
+                carry = (carry && (++j2 >= nr2));
+                j2 -= carry * nr2;
+                j3 += carry;
+            }
             dst_dd[i] = acc;
         });
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl

@@ -28,8 +28,11 @@ layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
 #endif
 
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+layout (binding = 3) readonly buffer Bias {D_TYPE data_bias[];};
+
 #ifdef MUL_MAT_ID
-layout (binding = 3) readonly buffer IDS {int data_ids[];};
+layout (binding = 4) readonly buffer IDS {int data_ids[];};
 #endif
 
 #include "dequant_funcs.glsl"
@@ -45,6 +48,8 @@ layout (push_constant) uniform parameter
     uint batch_stride_b;
     uint batch_stride_d;
 
+    uint enable_bias;
+
 #ifdef MUL_MAT_ID
     uint nei0;
     uint ne11;
@@ -56,6 +61,10 @@ layout (push_constant) uniform parameter
 #endif
 } p;
 
+#ifdef MUL_MAT_ID
+uint expert_id;
+#endif
+
 void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #ifdef MUL_MAT_ID
     const uint expert_idx = gl_GlobalInvocationID.y;
@@ -75,7 +84,7 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
         batch_idx_a = i03 * p.ne02 + i02;
     }
 #else
-    const uint expert_id = data_ids[expert_idx];
+    expert_id = data_ids[expert_idx];
 #endif
 
     a_offset =
@@ -113,6 +122,13 @@ void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t
     if (tid == 0) {
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                if (p.enable_bias != 0) {
+#ifdef MUL_MAT_ID
+                    temp[j][n] += FLOAT_TYPE(data_bias[expert_id*p.stride_d + first_row + n]);
+#else
+                    temp[j][n] += FLOAT_TYPE(data_bias[j*p.batch_stride_d + d_offset + first_row + n]);
+#endif
+                }
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
             }
         }
@@ -148,6 +164,13 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                 [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
                     temp[j][n] += tmpsh[j][n][s];
                 }
+                if (p.enable_bias != 0) {
+#ifdef MUL_MAT_ID
+                    temp[j][n] += FLOAT_TYPE(data_bias[expert_id*p.stride_d + first_row + n]);
+#else
+                    temp[j][n] += FLOAT_TYPE(data_bias[j*p.batch_stride_d + d_offset + first_row + n]);
+#endif
+                }
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
             }
         }
@@ -173,6 +196,13 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
     if (tid == 0) {
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                if (p.enable_bias != 0) {
+#ifdef MUL_MAT_ID
+                    tmpsh[j][n][0] += FLOAT_TYPE(data_bias[expert_id*p.stride_d + first_row + n]);
+#else
+                    tmpsh[j][n][0] += FLOAT_TYPE(data_bias[j*p.batch_stride_d + d_offset + first_row + n]);
+#endif
+                }
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][0]);
             }
         }

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

@@ -15,6 +15,8 @@ layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
 layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
 layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
 
+layout (binding = 3) readonly buffer Bias {D_TYPE data_bias[];};
+
 layout (push_constant) uniform parameter
 {
     uint ncols_x;
@@ -29,6 +31,7 @@ layout (push_constant) uniform parameter
     uint nb03;
     uint nb13;
     uint nb23;
+    uint enable_bias;
 } p;
 
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
@@ -117,6 +120,9 @@ void main() {
     }
 
     if (tid == 0) {
+        if (p.enable_bias != 0) {
+            tmp[0] += FLOAT_TYPE(data_bias[idst]);
+        }
         dst[idst] = tmp[0];
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp

@@ -17,6 +17,8 @@ layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
 layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
 layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
 
+layout (binding = 3) readonly buffer Bias {D_TYPE data_bias[];};
+
 layout(constant_id = 0) const int BLOCK_SIZE = 32;
 // gqa_ratio is in the range [1,8]
 layout(constant_id = 1) const uint gqa_ratio = 1;
@@ -29,6 +31,7 @@ layout (push_constant) uniform parameter
     uint nchannels_y;
     uint b_offset;
     uint d_offset;
+    uint enable_bias;
 } p;
 
 #if !USE_SUBGROUP_ADD
@@ -148,6 +151,9 @@ void main() {
         [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) {
             // dst is not transposed and not permuted
             const uint idst = (channel + c)*nrows_dst + row_dst;
+            if (p.enable_bias != 0) {
+                temp[c] += FLOAT_TYPE(data_bias[idst]);
+            }
             dst[idst] = temp[c];
         }
     }

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

@@ -23,16 +23,100 @@ layout (push_constant) uniform parameter2
     uint rms_partials;
 } p;
 
-// Workaround for MoltenVK Bug, see https://github.com/ggml-org/llama.cpp/issues/15498
-// layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
-// layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
-layout (binding = 0) buffer A {A_TYPE data_a[];} a[];
-layout (binding = 0) buffer D {D_TYPE data_d[];} d[];
-
-layout (binding = 0, std430) buffer PartialBuf {float partial_sums[];} partials[];
+// No readonly/writeonly decorations. Workaround for MoltenVK Bug, see https://github.com/ggml-org/llama.cpp/issues/15498
+layout (binding = 0)  buffer A0 {A_TYPE data_a[];} a0;
+layout (binding = 1)  buffer A1 {A_TYPE data_a[];} a1;
+layout (binding = 2)  buffer A2 {A_TYPE data_a[];} a2;
+layout (binding = 3)  buffer A3 {A_TYPE data_a[];} a3;
+layout (binding = 4)  buffer A4 {A_TYPE data_a[];} a4;
+layout (binding = 5)  buffer A5 {A_TYPE data_a[];} a5;
+layout (binding = 6)  buffer A6 {A_TYPE data_a[];} a6;
+layout (binding = 7)  buffer A7 {A_TYPE data_a[];} a7;
+layout (binding = 8)  buffer A8 {A_TYPE data_a[];} a8;
+layout (binding = 9)  buffer A9 {A_TYPE data_a[];} a9;
+layout (binding = 10) buffer A10 {A_TYPE data_a[];} a10;
+layout (binding = 11) buffer A11 {A_TYPE data_a[];} a11;
+layout (binding = 0)  buffer D0 {D_TYPE data_d[];} d0;
+layout (binding = 1)  buffer D1 {D_TYPE data_d[];} d1;
+layout (binding = 2)  buffer D2 {D_TYPE data_d[];} d2;
+layout (binding = 3)  buffer D3 {D_TYPE data_d[];} d3;
+layout (binding = 4)  buffer D4 {D_TYPE data_d[];} d4;
+layout (binding = 5)  buffer D5 {D_TYPE data_d[];} d5;
+layout (binding = 6)  buffer D6 {D_TYPE data_d[];} d6;
+layout (binding = 7)  buffer D7 {D_TYPE data_d[];} d7;
+layout (binding = 8)  buffer D8 {D_TYPE data_d[];} d8;
+layout (binding = 9)  buffer D9 {D_TYPE data_d[];} d9;
+layout (binding = 10) buffer D10 {D_TYPE data_d[];} d10;
+layout (binding = 11) buffer D11 {D_TYPE data_d[];} d11;
+layout (binding = 0, std430)  buffer PartialBuf0 {float partial_sums[];} partials0;
+layout (binding = 1, std430)  buffer PartialBuf1 {float partial_sums[];} partials1;
+layout (binding = 2, std430)  buffer PartialBuf2 {float partial_sums[];} partials2;
+layout (binding = 3, std430)  buffer PartialBuf3 {float partial_sums[];} partials3;
+layout (binding = 4, std430)  buffer PartialBuf4 {float partial_sums[];} partials4;
+layout (binding = 5, std430)  buffer PartialBuf5 {float partial_sums[];} partials5;
+layout (binding = 6, std430)  buffer PartialBuf6 {float partial_sums[];} partials6;
+layout (binding = 7, std430)  buffer PartialBuf7 {float partial_sums[];} partials7;
+layout (binding = 8, std430)  buffer PartialBuf8 {float partial_sums[];} partials8;
+layout (binding = 9, std430)  buffer PartialBuf9 {float partial_sums[];} partials9;
+layout (binding = 10, std430) buffer PartialBuf10 {float partial_sums[];} partials10;
+layout (binding = 11, std430) buffer PartialBuf11 {float partial_sums[];} partials11;
 
 layout(constant_id = 0) const uint num_srcs = 2;
 
+FLOAT_TYPE load_a(uint b, uint i) {
+    switch (b) {
+    case 0:  return FLOAT_TYPE(a0.data_a[i]);
+    case 1:  return FLOAT_TYPE(a1.data_a[i]);
+    case 2:  return FLOAT_TYPE(a2.data_a[i]);
+    case 3:  return FLOAT_TYPE(a3.data_a[i]);
+    case 4:  return FLOAT_TYPE(a4.data_a[i]);
+    case 5:  return FLOAT_TYPE(a5.data_a[i]);
+    case 6:  return FLOAT_TYPE(a6.data_a[i]);
+    case 7:  return FLOAT_TYPE(a7.data_a[i]);
+    case 8:  return FLOAT_TYPE(a8.data_a[i]);
+    case 9:  return FLOAT_TYPE(a9.data_a[i]);
+    case 10: return FLOAT_TYPE(a10.data_a[i]);
+    case 11: return FLOAT_TYPE(a11.data_a[i]);
+    default: return FLOAT_TYPE(0);
+    }
+}
+
+void store_d(uint b, uint i, FLOAT_TYPE v) {
+    switch (b) {
+    case 0:  d0.data_d[i] = D_TYPE(v); break;
+    case 1:  d1.data_d[i] = D_TYPE(v); break;
+    case 2:  d2.data_d[i] = D_TYPE(v); break;
+    case 3:  d3.data_d[i] = D_TYPE(v); break;
+    case 4:  d4.data_d[i] = D_TYPE(v); break;
+    case 5:  d5.data_d[i] = D_TYPE(v); break;
+    case 6:  d6.data_d[i] = D_TYPE(v); break;
+    case 7:  d7.data_d[i] = D_TYPE(v); break;
+    case 8:  d8.data_d[i] = D_TYPE(v); break;
+    case 9:  d9.data_d[i] = D_TYPE(v); break;
+    case 10: d10.data_d[i] = D_TYPE(v); break;
+    case 11: d11.data_d[i] = D_TYPE(v); break;
+    default: break;
+    }
+}
+
+void store_partial(uint b, uint i, float v) {
+    switch (b) {
+    case 0:  partials0.partial_sums[i] = v; break;
+    case 1:  partials1.partial_sums[i] = v; break;
+    case 2:  partials2.partial_sums[i] = v; break;
+    case 3:  partials3.partial_sums[i] = v; break;
+    case 4:  partials4.partial_sums[i] = v; break;
+    case 5:  partials5.partial_sums[i] = v; break;
+    case 6:  partials6.partial_sums[i] = v; break;
+    case 7:  partials7.partial_sums[i] = v; break;
+    case 8:  partials8.partial_sums[i] = v; break;
+    case 9:  partials9.partial_sums[i] = v; break;
+    case 10: partials10.partial_sums[i] = v; break;
+    case 11: partials11.partial_sums[i] = v; break;
+    default: break;
+    }
+}
+
 uint src_idx(uint s, uint i00, uint i01, uint i02, uint i03) {
     return i03*p.nb[s][3] + i02*p.nb[s][2] + i01*p.nb[s][1] + i00*p.nb[s][0];
 }
@@ -78,10 +162,10 @@ void main() {
 
         FLOAT_TYPE sum = FLOAT_TYPE(0);
         [[unroll]] for (uint s = 0; s < num_srcs; ++s) {
-            sum += FLOAT_TYPE(a[s].data_a[src_idx(s, i00, i01, i02, i03)]);
+            sum += load_a(s, src_idx(s, i00, i01, i02, i03));
         }
         sum_sq += sum*sum;
-        d[num_srcs].data_d[dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
+        store_d(num_srcs, dst_idx(i00, i01, i02, i03), sum);
 
         idx += num_threads;
     }
@@ -104,7 +188,7 @@ void main() {
         }
 
         if (gl_SubgroupID == 0 && gl_SubgroupInvocationID == 0) {
-            partials[num_srcs + 1].partial_sums[orig_idx / (num_iter * num_threads)] = sum_sq;
+            store_partial(num_srcs + 1, orig_idx / (num_iter * num_threads), sum_sq);
         }
     }
 #endif

gguf-py/gguf/constants.py

@@ -425,6 +425,7 @@ class MODEL_ARCH(IntEnum):
     GROVEMOE         = auto()
     APERTUS          = auto()
     COGVLM           = auto()
+    MINIMAXM2        = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -790,6 +791,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.SEED_OSS:         "seed_oss",
     MODEL_ARCH.GROVEMOE:         "grovemoe",
     MODEL_ARCH.APERTUS:          "apertus",
+    MODEL_ARCH.MINIMAXM2:        "minimax-m2",
     MODEL_ARCH.COGVLM:           "cogvlm",
 }
 
@@ -2921,6 +2923,24 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_CHEXP,
         MODEL_TENSOR.FFN_UP_CHEXP,
     ],
+    MODEL_ARCH.MINIMAXM2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+    ],
     MODEL_ARCH.COGVLM: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -3166,6 +3186,7 @@ class VisionProjectorType:
     KIMIVL = "kimivl"
     LIGHTONOCR = "lightonocr"
     COGVLM = "cogvlm"
+    JANUS_PRO = "janus_pro"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/tensor_mapping.py

@@ -381,6 +381,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.moe_statics.e_score_correction",        # ernie4.5-moe
             "model.layers.{bid}.mlp.gate.expert_bias",                      # bailingmoe2
             "model.layers.{bid}.feed_forward.expert_bias",                  # lfm2moe
+            "model.layers.{bid}.block_sparse_moe.e_score_correction",       # minimax-m2
         ),
 
         # Feed-forward up
@@ -1182,6 +1183,7 @@ class TensorNameMap:
             "model.mm_projector.mlp.mlp.{bid}",
             "vision_model.vision_adapter.mlp.fc{bid}", # llama 4
             "mlp1.{bid}", # InternVL
+            "model.aligner.fc1.hidden_layers.{bid}", # Janus Pro
         ),
 
         MODEL_TENSOR.V_MMPROJ_PEG: (
@@ -1290,6 +1292,7 @@ class TensorNameMap:
             "model.vision_tower.encoder.layer.{bid}.attention.projection_layer", # Intern-S1
             "vpm.encoder.layers.{bid}.self_attn.out_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM
+            "model.vision_model.encoder.layers.{bid}.self_attn.projection_layer", # Janus Pro
             "vision_model.model.layers.{bid}.self_attn.o_proj", # llama4
             "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral-hf
             "vision_encoder.transformer.layers.{bid}.attention.wo", # pixtral

include/llama.h

@@ -461,7 +461,10 @@ extern "C" {
     LLAMA_API bool llama_supports_gpu_offload(void);
     LLAMA_API bool llama_supports_rpc        (void);
 
+    // NOTE: After creating a llama_context, it is recommended to query the actual values using these functions
+    //       In some cases the requested values via llama_context_params may differ from the actual values used by the context
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_ctx_seq  (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);
@@ -585,7 +588,7 @@ extern "C" {
     LLAMA_API int32_t llama_adapter_meta_val_str_by_index(const struct llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size);
 
     // Manually free a LoRA adapter
-    // Note: loaded adapters will be free when the associated model is deleted
+    // NOTE: loaded adapters will be free when the associated model is deleted
     LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);
 
     // Get the invocation tokens if the current lora is an alora
@@ -1111,8 +1114,6 @@ extern "C" {
     //        // sample from the logits of the last token in the batch
     //        const llama_token id = llama_sampler_sample(smpl, ctx, -1);
     //
-    //        // accepting the token updates the internal state of certain samplers (e.g. grammar, repetition, etc.)
-    //        llama_sampler_accept(smpl, id);
     //        ...
     //    }
     //

scripts/bench-models.sh

@@ -0,0 +1,74 @@
+#!/usr/bin/env bash
+
+RESULTS="bench-models-results.txt"
+: > "$RESULTS"
+
+ARGS_BB="-c 270336 -npp 512,4096,8192 -npl 1,2,4,8,16,32 -ntg 32"
+ARGS_B="-d 0,4096,8192,16384,32768 -p 2048 -n 32"
+
+QUICK=0
+while (( "$#" )); do
+  case "$1" in
+    --quick) QUICK=1; shift ;;
+    *) shift ;;
+  esac
+done
+
+if (( QUICK )); then
+  ARGS_BB="-c 20480 -npp 512,4096 -npl 1,2,4 -ntg 32"
+  ARGS_B="-d 0 -p 2048 -n 32"
+fi
+
+run_model() {
+  local HFR=$1
+  local HFF=$2
+
+  printf "## ${HFR}\n" | tee -a "$RESULTS"
+  printf "\n" | tee -a "$RESULTS"
+  printf "Model: https://huggingface.co/${HFR}\n" | tee -a "$RESULTS"
+  printf "\n" | tee -a "$RESULTS"
+
+  printf -- "- \`llama-batched-bench\`\n" | tee -a "$RESULTS"
+  printf "\n" | tee -a "$RESULTS"
+
+  ./bin/llama-batched-bench \
+    -hfr "${HFR}" -hff "${HFF}" \
+    -m "${HFF}" -fa 1 -ub 2048 --no-mmap \
+    ${ARGS_BB} | tee -a "$RESULTS"
+
+  printf "\n" | tee -a "$RESULTS"
+
+  printf -- "- \`llama-bench\`\n" | tee -a "$RESULTS"
+  printf "\n" | tee -a "$RESULTS"
+
+  ./bin/llama-bench \
+    -m "${HFF}" -fa 1 -ub 2048 -mmp 0 \
+    ${ARGS_B} | tee -a "$RESULTS"
+
+  printf "\n" | tee -a "$RESULTS"
+
+  printf "\n"
+}
+
+run_model "ggml-org/gpt-oss-20b-GGUF"                       "gpt-oss-20b-mxfp4.gguf"
+run_model "ggml-org/gpt-oss-120b-GGUF"                      "gpt-oss-120b-mxfp4-00001-of-00003.gguf"
+run_model "ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF" "qwen3-coder-30b-a3b-instruct-q8_0.gguf"
+run_model "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF"             "qwen2.5-coder-7b-q8_0.gguf"
+run_model "ggml-org/gemma-3-4b-it-qat-GGUF"                 "gemma-3-4b-it-qat-Q4_0.gguf"
+
+if [[ -f models-extra.txt ]]; then
+    while read -r HFR HFF; do
+        [[ -z "$HFR" ]] && continue
+        run_model "$HFR" "$HFF"
+    done < models-extra.txt
+fi
+
+printf "\n=====================================\n"
+printf "\n"
+
+cat "$RESULTS"
+
+printf "\n"
+printf "Done! Results are written to $RESULTS\n"
+printf "\n"
+

scripts/sync-ggml.last

@@ -1 +1 @@
-72632094336524a9c809e129e8b1c52154543a5a
+e02fb860ccbba8967905bceff23b677e88105280

src/CMakeLists.txt

@@ -35,6 +35,100 @@ add_library(llama
             unicode-data.cpp
             unicode.cpp
             unicode.h
+            models/apertus.cpp
+            models/arcee.cpp
+            models/arctic.cpp
+            models/arwkv7.cpp
+            models/baichuan.cpp
+            models/bailingmoe.cpp
+            models/bailingmoe2.cpp
+            models/bert.cpp
+            models/bitnet.cpp
+            models/bloom.cpp
+            models/chameleon.cpp
+            models/chatglm.cpp
+            models/codeshell.cpp
+            models/cogvlm.cpp
+            models/cohere2-iswa.cpp
+            models/command-r.cpp
+            models/dbrx.cpp
+            models/deci.cpp
+            models/deepseek.cpp
+            models/deepseek2.cpp
+            models/dots1.cpp
+            models/dream.cpp
+            models/ernie4-5-moe.cpp
+            models/ernie4-5.cpp
+            models/exaone.cpp
+            models/exaone4.cpp
+            models/falcon-h1.cpp
+            models/falcon.cpp
+            models/gemma-embedding.cpp
+            models/gemma.cpp
+            models/gemma2-iswa.cpp
+            models/gemma3-iswa.cpp
+            models/gemma3n-iswa.cpp
+            models/glm4-moe.cpp
+            models/glm4.cpp
+            models/gpt2.cpp
+            models/gptneox.cpp
+            models/granite-hybrid.cpp
+            models/granite.cpp
+            models/grok.cpp
+            models/grovemoe.cpp
+            models/hunyuan-dense.cpp
+            models/hunyuan-moe.cpp
+            models/internlm2.cpp
+            models/jais.cpp
+            models/jamba.cpp
+            models/lfm2.cpp
+            models/llada-moe.cpp
+            models/llada.cpp
+            models/llama-iswa.cpp
+            models/llama.cpp
+            models/mamba.cpp
+            models/minicpm3.cpp
+            models/minimax-m2.cpp
+            models/mpt.cpp
+            models/nemotron-h.cpp
+            models/nemotron.cpp
+            models/neo-bert.cpp
+            models/olmo.cpp
+            models/olmo2.cpp
+            models/olmoe.cpp
+            models/openai-moe-iswa.cpp
+            models/openelm.cpp
+            models/orion.cpp
+            models/phi2.cpp
+            models/phi3.cpp
+            models/plamo.cpp
+            models/plamo2.cpp
+            models/plm.cpp
+            models/qwen.cpp
+            models/qwen2.cpp
+            models/qwen2moe.cpp
+            models/qwen2vl.cpp
+            models/qwen3.cpp
+            models/qwen3vl.cpp
+            models/qwen3vl-moe.cpp
+            models/qwen3moe.cpp
+            models/refact.cpp
+            models/rwkv6-base.cpp
+            models/rwkv6.cpp
+            models/rwkv6qwen2.cpp
+            models/rwkv7-base.cpp
+            models/rwkv7.cpp
+            models/seed-oss.cpp
+            models/smallthinker.cpp
+            models/smollm3.cpp
+            models/stablelm.cpp
+            models/starcoder.cpp
+            models/starcoder2.cpp
+            models/t5-dec.cpp
+            models/t5-enc.cpp
+            models/wavtokenizer-dec.cpp
+            models/xverse.cpp
+            models/graph-context-mamba.cpp
             )
 
 target_include_directories(llama PRIVATE .)

src/llama-arch.cpp

@@ -105,6 +105,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_SEED_OSS,         "seed_oss"         },
     { LLM_ARCH_GROVEMOE,         "grovemoe"         },
     { LLM_ARCH_APERTUS,          "apertus"          },
+    { LLM_ARCH_MINIMAX_M2,       "minimax-m2"       },
     { LLM_ARCH_COGVLM,           "cogvlm"           },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
@@ -2355,6 +2356,27 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_CHEXPS,      "blk.%d.ffn_up_chexps" },
         },
     },
+    {
+        LLM_ARCH_MINIMAX_M2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+        },
+    },
     {
         LLM_ARCH_COGVLM,
         {

src/llama-arch.h

@@ -109,6 +109,7 @@ enum llm_arch {
     LLM_ARCH_SEED_OSS,
     LLM_ARCH_GROVEMOE,
     LLM_ARCH_APERTUS,
+    LLM_ARCH_MINIMAX_M2,
     LLM_ARCH_COGVLM,
     LLM_ARCH_UNKNOWN,
 };

src/llama-context.cpp

@@ -112,11 +112,24 @@ llama_context::llama_context(
         }
     }
 
-    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
+    if (cparams.kv_unified) {
+        cparams.n_ctx_seq = cparams.n_ctx;
+    } else {
+        cparams.n_ctx_seq = cparams.n_ctx / cparams.n_seq_max;
+
+        if (cparams.n_ctx_seq == 0) {
+            throw std::runtime_error("n_ctx_seq == 0");
+        }
+
+        if (cparams.n_ctx != cparams.n_ctx_seq * cparams.n_seq_max) {
+            cparams.n_ctx =  cparams.n_ctx_seq * cparams.n_seq_max;
+            LLAMA_LOG_WARN("%s: n_ctx is not divisible by n_seq_max - rounding down to %u\n", __func__, cparams.n_ctx);
+        }
+    }
 
     LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",   __func__, cparams.n_seq_max);
     LLAMA_LOG_INFO("%s: n_ctx         = %u\n",   __func__, cparams.n_ctx);
-    LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n",   __func__, n_ctx_per_seq);
+    LLAMA_LOG_INFO("%s: n_ctx_seq     = %u\n",   __func__, cparams.n_ctx_seq);
     LLAMA_LOG_INFO("%s: n_batch       = %u\n",   __func__, cparams.n_batch);
     LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
     LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
@@ -125,14 +138,14 @@ llama_context::llama_context(
     LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
     LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
 
-    if (n_ctx_per_seq < hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    if (cparams.n_ctx_seq < hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
+                __func__, cparams.n_ctx_seq, hparams.n_ctx_train);
     }
 
-    if (n_ctx_per_seq > hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    if (cparams.n_ctx_seq > hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
+                __func__, cparams.n_ctx_seq, hparams.n_ctx_train);
     }
 
     if (!hparams.vocab_only) {
@@ -453,8 +466,8 @@ uint32_t llama_context::n_ctx() const {
     return cparams.n_ctx;
 }
 
-uint32_t llama_context::n_ctx_per_seq() const {
-    return cparams.n_ctx / cparams.n_seq_max;
+uint32_t llama_context::n_ctx_seq() const {
+    return cparams.n_ctx_seq;
 }
 
 uint32_t llama_context::n_batch() const {
@@ -2383,6 +2396,10 @@ uint32_t llama_n_ctx(const llama_context * ctx) {
     return ctx->n_ctx();
 }
 
+uint32_t llama_n_ctx_seq(const llama_context * ctx) {
+    return ctx->n_ctx_seq();
+}
+
 uint32_t llama_n_batch(const llama_context * ctx) {
     return ctx->n_batch();
 }

src/llama-context.h

@@ -43,11 +43,11 @@ struct llama_context {
 
     ggml_backend_sched_t get_sched() const;
 
-    uint32_t n_ctx()         const;
-    uint32_t n_ctx_per_seq() const;
-    uint32_t n_batch()       const;
-    uint32_t n_ubatch()      const;
-    uint32_t n_seq_max()     const;
+    uint32_t n_ctx()     const;
+    uint32_t n_ctx_seq() const;
+    uint32_t n_batch()   const;
+    uint32_t n_ubatch()  const;
+    uint32_t n_seq_max() const;
 
     uint32_t n_threads()       const;
     uint32_t n_threads_batch() const;

src/llama-cparams.h

@@ -8,6 +8,7 @@
 
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference
+    uint32_t n_ctx_seq;       // context for a single sequence
     uint32_t n_batch;
     uint32_t n_ubatch;
     uint32_t n_seq_max;

src/llama-model.h

@@ -114,6 +114,7 @@ enum llm_type {
     LLM_TYPE_30B_A3B,
     LLM_TYPE_100B_A6B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air
+    LLM_TYPE_230B_A10B, // Minimax M2
     LLM_TYPE_235B_A22B,
     LLM_TYPE_300B_A47B, // Ernie MoE big
     LLM_TYPE_355B_A32B, // GLM-4.5

src/llama-vocab.cpp

@@ -401,6 +401,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                 };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_GPT4O:
+            case LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2:
                 regex_exprs = {
                     // original regex from tokenizer.json
                     // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -1992,6 +1993,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "grok-2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GROK_2;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "minimax-m2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;
+                clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }

src/llama-vocab.h

@@ -49,6 +49,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE   = 38,
     LLAMA_VOCAB_PRE_TYPE_GROK_2          = 39,
     LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40,
+    LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2      = 41,
 };
 
 struct LLM_KV;

src/models/apertus.cpp

@@ -0,0 +1,125 @@
+#include "models.h"
+
+
+
+llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos  = build_inp_pos();
+    auto *        inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+            cb(Vcur, "Vcur_pos", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network with xIELU activation
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // Up projection
+            ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur);
+            cb(up, "ffn_up", il);
+
+            float alpha_n_val = hparams.xielu_alpha_n[il];
+            float alpha_p_val = hparams.xielu_alpha_p[il];
+            float beta_val    = hparams.xielu_beta[il];
+            float eps_val     = hparams.xielu_eps[il];
+
+            // Apply xIELU activation
+            ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val);
+            cb(activated, "ffn_xielu", il);
+
+            // Down projection
+            cur = build_lora_mm(model.layers[il].ffn_down, activated);
+            cb(cur, "ffn_down", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/arcee.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        // ARCEE uses relu^2 instead of silu
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                NULL,                      NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/arctic.cpp

@@ -0,0 +1,138 @@
+#include "models.h"
+
+
+llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
+        cb(ffn_out, "ffn_out", il);
+
+        // MoE
+        cur = build_norm(inpSA,
+                model.layers[il].ffn_norm_exps, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm_exps", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_out);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/arwkv7.cpp

@@ -0,0 +1,86 @@
+#include "models.h"
+
+
+llm_build_arwkv7::llm_build_arwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
+    GGML_ASSERT(n_embd == hparams.n_embd_r());
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * v_first = nullptr;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * rs_inp = build_rs_inp();
+
+    const auto n_embd = hparams.n_embd;
+    const auto n_seq_tokens = ubatch.n_seq_tokens;
+    const auto n_seqs = ubatch.n_seqs;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const llama_layer * layer = &model.layers[il];
+        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
+
+        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
+
+        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
+        cb(att_norm, "attn_norm", il);
+
+        ggml_tensor * x_prev = ggml_concat(
+                ctx0,
+                token_shift,
+                ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
+                1
+                );
+
+        cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
+
+        token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
+        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
+        ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/baichuan.cpp

@@ -0,0 +1,122 @@
+#include "models.h"
+
+
+llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            switch (model.type) {
+                case LLM_TYPE_7B:
+                    Qcur = ggml_rope_ext(
+                            ctx0, Qcur, inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                    Kcur = ggml_rope_ext(
+                            ctx0, Kcur, inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                    break;
+                case LLM_TYPE_13B:
+                    break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/bailingmoe.cpp

@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, hparams.expert_weights_norm,
+                    false, hparams.expert_weights_scale,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // FFN shared expert
+        {
+            ggml_tensor * ffn_shexp = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/bailingmoe2.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+
+llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+    for (int il = 0; il < n_transformer_layers; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 0 * sizeof(float) * (n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 1 * sizeof(float) * (n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_transformer_layers - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA);
+        cb(sa_out, "sa_out", il);
+
+        // MoE branch
+        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/bert.cpp

@@ -0,0 +1,176 @@
+#include "models.h"
+
+
+
+llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * inp_pos = nullptr;
+
+    if (model.arch != LLM_ARCH_JINA_BERT_V2) {
+        inp_pos = build_inp_pos();
+    }
+
+    // construct input embeddings (token, type, position)
+    inpL = build_inp_embd(model.tok_embd);
+
+    // token types are hardcoded to zero ("Sentence A")
+    if (model.type_embd) {
+        ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
+        inpL                    = ggml_add(ctx0, inpL, type_row0);
+    }
+    if (model.arch == LLM_ARCH_BERT) {
+        inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
+    }
+    cb(inpL, "inp_embd", -1);
+
+    // embed layer norm
+    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
+    cb(inpL, "inp_norm", -1);
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * cur = inpL;
+
+        {
+            ggml_tensor * Qcur;
+            ggml_tensor * Kcur;
+            ggml_tensor * Vcur;
+
+            // self-attention
+            if (model.layers[il].wqkv) {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
+                                    0 * sizeof(float) * (n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+            } else {
+                Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq);
+                Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk);
+                Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            }
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);
+
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            }
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);
+
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);
+
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            }
+
+            // RoPE
+            if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
+                model.arch == LLM_ARCH_JINA_BERT_V3) {
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "kqv_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // re-add the layer input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        // attention layer norm
+        cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il);
+
+        if (model.layers[il].attn_norm_2 != nullptr) {
+            cur = ggml_add(ctx0, cur, inpL);  // re-add the layer input
+            cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il);
+        }
+
+        ggml_tensor * ffn_inp = cur;
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
+            // MoE branch
+            cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, nullptr,
+                                model.layers[il].ffn_down_exps, nullptr, hparams.n_expert, hparams.n_expert_used,
+                                LLM_FFN_GELU, false, false, 0.0f, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
+            cb(cur, "ffn_moe_out", il);
+        } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
+                   model.arch == LLM_ARCH_JINA_BERT_V3) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                    NULL, NULL, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
+                    model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // attentions bypass the intermediate layer
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        // output layer norm
+        cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cb(cur, "result_embd", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/bitnet.cpp

@@ -0,0 +1,160 @@
+#include "models.h"
+
+
+llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            if (model.layers[il].wq_scale) {
+                Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale);
+            }
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            // B1.K
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            if (model.layers[il].wk_scale) {
+                Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale);
+            }
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            // B1.V
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            if (model.layers[il].wv_scale) {
+                Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale);
+            }
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    NULL, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+
+            cur = build_norm(cur,
+                    model.layers[il].attn_sub_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_sub_norm", il);
+
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            if (model.layers[il].wo_scale) {
+                cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale);
+            }
+            if (model.layers[il].bo) {
+                cur = ggml_add(ctx0, cur, model.layers[il].bo);
+            }
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward forward
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_scale,
+                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale,
+                NULL,                      NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_sub_out", il);
+
+        cur = build_norm(cur,
+                model.layers[il].ffn_sub_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_sub_norm", il);
+
+        cur = build_lora_mm(model.layers[il].ffn_down, cur);
+        if (model.layers[il].ffn_down_scale) {
+            cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale);
+        }
+        cb(cur, "ffn_down", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    // FIXME: do not use model.tok_embd directly, duplicate as model.output
+    cur = build_lora_mm(model.tok_embd, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/bloom.cpp

@@ -0,0 +1,101 @@
+#include "models.h"
+
+llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    inpL = build_norm(inpL,
+            model.tok_norm,
+            model.tok_norm_b,
+            LLM_NORM, -1);
+    cb(inpL, "inp_norm", -1);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // Add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/chameleon.cpp

@@ -0,0 +1,178 @@
+#include "models.h"
+
+#include <float.h>
+
+llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        if (hparams.swin_norm) {
+            cur = inpL;
+        } else {
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
+                        ggml_element_size(Qcur) * n_embd_head,
+                        ggml_element_size(Qcur) * n_embd_head * n_head,
+                        0);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = build_norm(Qcur,
+                        model.layers[il].attn_q_norm,
+                        model.layers[il].attn_q_norm_b,
+                        LLM_NORM, il);
+                cb(Qcur, "Qcur", il);
+            }
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
+                        ggml_element_size(Kcur) * n_embd_head,
+                        ggml_element_size(Kcur) * n_embd_head * n_head_kv,
+                        0);
+                cb(Kcur, "Kcur", il);
+
+                Kcur = build_norm(Kcur,
+                        model.layers[il].attn_k_norm,
+                        model.layers[il].attn_k_norm_b,
+                        LLM_NORM, il);
+                cb(Kcur, "Kcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        if (hparams.swin_norm) {
+            cur = build_norm(cur,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        if (!hparams.swin_norm) {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+        }
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        if (hparams.swin_norm) {
+            cur = build_norm(cur,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output_with_img_logits", -1);
+
+    // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.
+    // Needs to be removed once image outputs are supported.
+    int img_token_end_idx = 8196;
+    int img_token_start_idx = 4;
+    int num_img_tokens = img_token_end_idx - img_token_start_idx;
+    // creates 1d tensor of size num_img_tokens and values -FLT_MAX,
+    // which ensures that text token values are always at least larger than image token values
+    ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);
+    img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX);
+    cb(img_logits, "img_logits", -1);
+
+    cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/chatglm.cpp

@@ -0,0 +1,132 @@
+#include "models.h"
+
+
+llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv == nullptr) {
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            } else {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+            }
+
+            //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        // Add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+        }
+
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        cb(inpL, "l_out", il);
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/codeshell.cpp

@@ -0,0 +1,111 @@
+#include "models.h"
+
+llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/cogvlm.cpp

@@ -0,0 +1,100 @@
+#include "models.h"
+
+llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    float         kq_scale    = 1.0f / sqrtf(float(n_embd_head));
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor *inpL, *cur;
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    // check ubatch to see if we have input tokens (text)
+    // or an input embedding vector (image)
+    bool is_text;
+    if (ubatch.token) {
+        is_text = true;
+    } else {
+        is_text = false;
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        // get either the text or image weight tensors
+        ggml_tensor *wqkv, *wo;
+        ggml_tensor *ffn_gate, *ffn_down, *ffn_up;
+
+        if (is_text) {
+            wqkv     = model.layers[il].wqkv;
+            wo       = model.layers[il].wo;
+            ffn_gate = model.layers[il].ffn_gate;
+            ffn_down = model.layers[il].ffn_down;
+            ffn_up   = model.layers[il].ffn_up;
+        } else {
+            wqkv     = model.layers[il].visexp_attn_wqkv;
+            wo       = model.layers[il].visexp_attn_wo;
+            ffn_gate = model.layers[il].visexp_ffn_gate;
+            ffn_down = model.layers[il].visexp_ffn_down;
+            ffn_up   = model.layers[il].visexp_ffn_up;
+        }
+
+        ggml_tensor * inpSA = inpL;
+        cur                 = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+
+        // build self attention
+        {
+            ggml_tensor * qkv = build_lora_mm(wqkv, cur);
+
+            // split qkv into Q, K, V along the first dimension
+            ggml_tensor * Qcur =
+                ggml_view_3d(ctx0, qkv, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), qkv->nb[1], 0);
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              qkv->nb[1], n_embd * ggml_element_size(qkv));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              qkv->nb[1], 2 * n_embd * ggml_element_size(qkv));
+
+            Qcur = ggml_rope(ctx0, Qcur, inp_pos, n_embd_head, rope_type);
+            Kcur = ggml_rope(ctx0, Kcur, inp_pos, n_embd_head, rope_type);
+
+            cur = build_attn(inp_attn,
+                wo, nullptr,
+                Qcur, Kcur, Vcur,
+                nullptr, nullptr, nullptr,
+                kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                ffn_up, NULL, NULL,
+                ffn_gate, NULL, NULL,
+                ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}

src/models/cohere2-iswa.cpp

@@ -0,0 +1,131 @@
+#include "models.h"
+
+llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    const float f_logit_scale = hparams.f_logit_scale;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const bool is_swa = hparams.is_swa(il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+        ggml_tensor * ffn_inp = cur;
+
+        // self-attention
+        {
+            // rope freq factors for 128k context
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (is_swa) {
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+
+        ggml_tensor * attn_out = cur;
+
+        // feed-forward network
+        {
+            cur = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // add together residual + FFN + self-attention
+        cur = ggml_add(ctx0, cur, inpL);
+        cur = ggml_add(ctx0, cur, attn_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, f_logit_scale);
+    }
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/command-r.cpp

@@ -0,0 +1,122 @@
+#include "models.h"
+
+
+
+llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    const float f_logit_scale = hparams.f_logit_scale;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        ggml_tensor * ffn_inp = cur;
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM, il);
+                cb(Qcur, "Qcur", il);
+            }
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM, il);
+                cb(Kcur, "Kcur", il);
+            }
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+        ggml_tensor * attn_out = cur;
+
+        // feed-forward network
+        {
+            cur = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        // add together residual + FFN + self-attention
+        cur = ggml_add(ctx0, cur, inpL);
+        cur = ggml_add(ctx0, cur, attn_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, f_logit_scale);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/dbrx.cpp

@@ -0,0 +1,123 @@
+#include "models.h"
+
+
+llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+            cb(cur, "wqkv_clamped", il);
+
+            Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].attn_out_norm, NULL,
+                LLM_NORM, il);
+        cb(cur, "attn_out_norm", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/deci.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+
+llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA     = inpL;
+        const int64_t n_head_kv = hparams.n_head_kv(il);
+        const int64_t n_head    = hparams.n_head(il);
+        const int64_t n_ff      = hparams.n_ff(il);
+
+        if (n_head == 0) {
+            // attention-free layer of Llama-3_1-Nemotron-51B
+            cur = inpL;
+        } else {
+            // norm
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        if (n_head > 0 && n_head_kv == 0) {
+            // "linear attention" of Llama-3_1-Nemotron-51B
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            cb(cur, "wo", il);
+        } else if (n_head > 0) {
+            // self-attention
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B
+        if (n_ff == 0) {
+            continue;
+        }
+        // modified to support attention-free layer of Llama-3_1-Nemotron-51B
+        ggml_tensor * ffn_inp = cur;
+        if (n_head > 0) {
+            ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+        }
+        // feed-forward network
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/deepseek.cpp

@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+
+llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, false,
+                false, hparams.expert_weights_scale,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/deepseek2.cpp

@@ -0,0 +1,236 @@
+#include "models.h"
+
+
+
+llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    bool is_lite = (hparams.n_layer == 27);
+
+    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+
+    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+    const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
+    const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+    const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+
+    const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+    // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
+    // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
+    const float mscale      = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+    const float kq_scale    = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));
+    const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // {n_embd, n_tokens}
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            ggml_tensor * q = NULL;
+            if (!is_lite) {
+                q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+                cb(q, "q", il);
+
+                q = build_norm(q, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+                cb(q, "q", il);
+
+                q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
+                cb(q, "q", il);
+            } else {
+                q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(q, "q", il);
+            }
+            // split into {n_embd_head_qk_nope, n_head, n_tokens}
+            ggml_tensor * q_nope =
+                ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+                             ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+            cb(q_nope, "q_nope", il);
+
+            // and {n_embd_head_qk_rope, n_head, n_tokens}
+            ggml_tensor * q_pe = ggml_view_3d(
+                ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+                ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope));
+            cb(q_pe, "q_pe", il);
+
+            ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+            cb(kv_cmpr_pe, "kv_cmpr_pe", il);
+
+            // split into {kv_lora_rank, n_tokens}
+            ggml_tensor * kv_cmpr =
+                ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+                             ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            // and {n_embd_head_qk_rope, 1, n_tokens}
+            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+            cb(k_pe, "k_pe", il);
+
+            q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(q_pe, "q_pe", il);
+
+            k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(k_pe, "k_pe", il);
+
+            kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            if (is_mla) {
+                // {n_embd_head_qk_nope, n_tokens, n_head}
+                q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+                cb(q_nope, "q_nope_perm", il);
+
+                // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
+                ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+                cb(q_nope_absorbed, "q_nope_absorbed", il);
+
+                // {kv_lora_rank, n_head, n_tokens}
+                q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+                cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
+                // note: rope must go first for in-place context shifting in build_rope_shift()
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
+                cb(Qcur, "Qcur", il);
+
+                kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+                cb(kv_cmpr, "kv_cmpr_reshape", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
+                cb(Kcur, "Kcur", il);
+
+                // {kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Vcur = kv_cmpr;
+                cb(Vcur, "Vcur", il);
+
+                // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
+            } else {
+                ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
+                cb(kv, "kv", il);
+
+                // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                ggml_tensor * k_nope =
+                    ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                                 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                                 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, 0);
+                cb(k_nope, "k_nope_view", il);
+
+                // and {n_embd_head_v, n_head, n_tokens}
+                ggml_tensor * Vcur = ggml_view_3d(ctx0, kv, n_embd_head_v, n_head, n_tokens,
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope));
+                cb(Vcur, "Vcur_view", il);
+
+                Vcur = ggml_cont(ctx0, Vcur);
+                cb(Vcur, "Vcur_cont", il);
+
+                // note: rope must go first for in-place context shifting in build_rope_shift()
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
+                cb(Kcur, "Kcur", il);
+
+                // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
+                cur = build_attn(inp_attn,
+                            model.layers[il].wo, NULL,
+                            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            }
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/dots1.cpp

@@ -0,0 +1,134 @@
+#include "models.h"
+
+
+
+llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/dream.cpp

@@ -0,0 +1,105 @@
+#include "models.h"
+
+
+
+llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    //copied from qwen2
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            Qcur               = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            Kcur               = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            Vcur               = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+            model.layers[il].ffn_up, NULL, NULL,
+            model.layers[il].ffn_gate, NULL, NULL,
+            model.layers[il].ffn_down, NULL, NULL,
+            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/ernie4-5-moe.cpp

@@ -0,0 +1,150 @@
+#include "models.h"
+
+
+
+llm_build_ernie4_5_moe::llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0");
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+        // norm
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        bool is_moe_layer =
+            static_cast<uint32_t>(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0;
+
+        if (!is_moe_layer) {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                                        model.layers[il].ffn_gate_inp,
+                                        model.layers[il].ffn_up_exps,
+                                        model.layers[il].ffn_gate_exps,
+                                        model.layers[il].ffn_down_exps,
+                                        model.layers[il].ffn_exp_probs_b,
+                                        n_expert, n_expert_used,
+                                        LLM_FFN_SILU, true,
+                                        false, 0.0,
+                                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                                        il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // Shared expert (if present)
+            if (hparams.n_ff_shexp > 0) {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            } else {
+                cur = moe_out;
+            }
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/ernie4-5.cpp

@@ -0,0 +1,111 @@
+#include "models.h"
+
+
+
+llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1) {
+            // skip computing output for unused tokens
+            ggml_tensor * inp_out_ids = build_inp_out_ids();
+            cur                       = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA                     = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/exaone.cpp

@@ -0,0 +1,114 @@
+#include "models.h"
+
+
+
+llm_build_exaone::llm_build_exaone(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/exaone4.cpp

@@ -0,0 +1,123 @@
+#include "models.h"
+
+
+template <bool iswa>
+llm_build_exaone4<iswa>::llm_build_exaone4(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    using inp_attn_type      = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // use RoPE for SWA layers or non-SWA models
+        const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE;
+
+        cur = inpL;
+
+        // self-attention
+        {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+
+            if (use_rope) {
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
+                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
+                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            }
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_ffn(ffn_inp,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL, NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+// Explicit template instantiations
+template struct llm_build_exaone4<false>;
+template struct llm_build_exaone4<true>;

src/models/falcon-h1.cpp

@@ -0,0 +1,113 @@
+#include "models.h"
+
+
+
+llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // Build the inputs in the recurrent & kv cache
+    auto * inp = build_inp_mem_hybrid();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+        cb(Qcur, "Qcur", il);
+
+        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+        cb(Kcur, "Kcur", il);
+
+        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+        cb(Vcur, "Vcur", il);
+
+        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        cb(Qcur, "Qcur-post-rope", il);
+        cb(Kcur, "Kcur-post-rope", il);
+        cb(Vcur, "Vcur-post-rope", il);
+
+        ggml_tensor * attn_out = build_attn(inp->get_attn(),
+                                    model.layers[il].wo, NULL,
+                                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        cb(attn_out, "attn_out", il);
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        // Mamba2 layer
+        cb(cur, "ssm_in", il);
+
+        ggml_tensor * ssm_out = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
+        cb(ssm_out, "ssm_out", il);
+
+        // // Aggregation
+        cur   = ggml_add(ctx0, attn_out, ssm_out);
+        inpSA = ggml_add(ctx0, cur, inpSA);
+        cb(cur, "layer_out", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = inpSA;
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, inpSA);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/falcon.cpp

@@ -0,0 +1,120 @@
+#include "models.h"
+
+
+llm_build_falcon::llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * attn_norm;
+
+        attn_norm = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(attn_norm, "attn_norm", il);
+
+        // self-attention
+        {
+            if (model.layers[il].attn_norm_2) {
+                // Falcon-40B
+                cur = build_norm(inpL,
+                        model.layers[il].attn_norm_2,
+                        model.layers[il].attn_norm_2_b,
+                        LLM_NORM, il);
+                cb(cur, "attn_norm_2", il);
+            } else {
+                cur = attn_norm;
+            }
+
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            // using mode = 2 for neox mode
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur       = ggml_get_rows(ctx0,       cur, inp_out_ids);
+            inpL      = ggml_get_rows(ctx0,      inpL, inp_out_ids);
+            attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = cur;
+
+        // feed forward
+        {
+            cur = build_ffn(attn_norm, // !! use the attn norm, not the result
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = ggml_add(ctx0, cur, inpL);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // norm
+    cur = build_norm(cur,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gemma-embedding.cpp

@@ -0,0 +1,120 @@
+#include "models.h"
+
+
+
+llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+            cur =
+                build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gemma.cpp

@@ -0,0 +1,112 @@
+#include "models.h"
+
+
+llm_build_gemma::llm_build_gemma(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+    cb(inpL, "inp_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
+            cb(Qcur, "Qcur_scaled", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gemma2-iswa.cpp

@@ -0,0 +1,125 @@
+#include "models.h"
+
+llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+    cb(inpL, "inp_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    // final logit soft-capping
+    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+    cur = ggml_tanh(ctx0, cur);
+    cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gemma3-iswa.cpp

@@ -0,0 +1,131 @@
+#include "models.h"
+
+llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // TODO: is causal == true correct? might need some changes
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        cur = build_norm(cur,
+                model.layers[il].attn_post_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = build_norm(cur,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gemma3n-iswa.cpp

@@ -0,0 +1,377 @@
+#include "models.h"
+
+
+
+llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params),
+    model(model),
+    n_embd_head(model.hparams.n_embd_head_k),
+    n_embd_altup(model.hparams.n_embd_altup),
+    n_altup(model.hparams.n_altup),
+    i_altup_act(model.hparams.i_altup_act) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // TODO: is causal == true correct? might need some changes
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer]
+    ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs());
+
+    // inpL now has only 1 altup, project it to the rest of the altups
+    // these "added" altups will be concat to the last dim of inpL
+    {
+        ggml_tensor * target_magnitude = calc_magnitude(inpL);
+        ggml_tensor * inp_repeated     = ggml_repeat_4d(ctx0, inpL, n_embd, n_tokens, n_altup - 1, 1);
+        ggml_tensor * altup_added =
+            ggml_mul_mat(ctx0, model.altup_proj, inp_repeated);  // shape: [n_embd, n_tokens, n_altup - 1]
+        ggml_tensor * new_magnitude = calc_magnitude(altup_added);
+        altup_added                 = ggml_div(ctx0, ggml_mul(ctx0, altup_added, target_magnitude), new_magnitude);
+        inpL                        = ggml_concat(ctx0, inpL, altup_added, 2);  // shape: [n_embd, n_tokens, n_altup]
+        cb(inpL, "inp_stacked", -1);
+    }
+    // inpL now has shape:          [n_embd,       n_tokens, n_altup]
+    // inp_per_layer now has shape: [n_embd_altup, n_tokens, n_layer]
+
+    for (int il = 0; il < n_layer; ++il) {
+        // this block is made to be closely resemble Gemma3p5DecoderLayer on python code
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        ggml_tensor * cur         = inpL;                    // [n_embd, n_tokens, n_altup]
+        ggml_tensor * predictions = altup_predict(cur, il);  // [n_embd, n_tokens, n_altup]
+
+        // predicted value will go through self-attention and laurel
+        ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);  // [n_embd, n_tokens]
+        cur                             = active_prediction;
+        cb(cur, "active_prediction", il);
+
+        // norm
+        cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // laurel
+        ggml_tensor * laurel_out = laurel(cur, il);  // [n_embd, n_tokens]
+
+        // self-attention
+        if (hparams.has_kv(il)) {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps);
+
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+            cb(Vcur, "Vcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+
+            cur = build_attn(inp_attn, model.layers[il].wo,
+                    NULL, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
+                    hparams.f_attention_scale, il);
+        } else {
+            // reuse KV cache of earlier layers
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(Qcur, "Qcur_pos", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
+        }
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, active_prediction);  // [n_embd, n_tokens]
+        cb(cur, "attn_gated", il);
+
+        ggml_tensor * attn_laurel = ggml_scale(ctx0, ggml_add(ctx0, cur, laurel_out),
+                                               1.0f / sqrtf(2.0f));  // [n_embd, n_tokens]
+        cb(attn_laurel, "attn_laurel", il);
+
+        cur = build_norm(attn_laurel, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            ggml_tensor * up_proj   = build_lora_mm(model.layers[il].ffn_up, cur);
+            ggml_tensor * gate_proj = build_lora_mm(model.layers[il].ffn_gate, cur);
+
+            if (il < n_layer_sparsity) {
+                // apply activation sparsity
+                gate_proj = gaussian_topk(gate_proj);
+            }
+            gate_proj = ggml_gelu(ctx0, gate_proj);
+
+            cur = ggml_mul(ctx0, up_proj, gate_proj);
+            cur = build_lora_mm(model.layers[il].ffn_down, cur);
+            cb(cur, "ffn_out", il);
+        }
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", il);
+
+        ggml_tensor * attn_ffw_laurel_gated = ggml_add(ctx0, cur, attn_laurel);  // [n_embd, n_tokens]
+        cb(attn_ffw_laurel_gated, "attn_ffw_laurel_gated", il);
+
+        ggml_tensor * corrected = altup_correct(predictions, attn_ffw_laurel_gated, il);  // [n_embd, n_tokens, n_altup]
+
+        ggml_tensor * first_prediction;                                                   // [n_embd, n_tokens]
+        {
+            first_prediction = view_2d_slice(corrected, i_altup_act);                     // [n_embd, n_tokens]
+            first_prediction = ggml_mul(ctx0, first_prediction, model.layers[il].altup_correct_scale);
+            first_prediction = build_lora_mm(model.layers[il].per_layer_inp_gate, first_prediction);
+            first_prediction = ggml_gelu(ctx0, first_prediction);                 // [n_embd_altup, n_tokens]
+            cb(first_prediction, "first_prediction_gated", il);
+            ggml_tensor * inp_this_layer = view_2d_slice(inp_per_layer, il);      // [n_embd_altup, n_tokens]
+            first_prediction = ggml_mul(ctx0, first_prediction, inp_this_layer);  // [n_embd_altup, n_tokens]
+            cb(first_prediction, "first_prediction_scaled", il);
+
+            first_prediction = build_lora_mm(model.layers[il].per_layer_proj, first_prediction);  // [n_embd, n_tokens]
+            first_prediction =
+                build_norm(first_prediction, model.layers[il].per_layer_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(first_prediction, "first_prediction_out", il);
+        }
+        // equivalent to python code: corrected_predictions[1:] += first_prediction
+        {
+            ggml_tensor * slice_first = view_2d_slice(corrected, 0);
+            ggml_tensor * slice_rest  = ggml_view_3d(
+                ctx0, corrected, n_embd, n_tokens, n_altup - 1, ggml_row_size(corrected->type, n_embd),
+                ggml_row_size(corrected->type, n_embd * n_tokens), n_embd * n_tokens * ggml_element_size(corrected));
+            ggml_tensor * tmp = ggml_add(ctx0, slice_rest, first_prediction);  // [n_embd, n_tokens, n_altup - 1]
+            corrected         = ggml_concat(ctx0, slice_first, tmp, 2);        // [n_embd, n_tokens, n_altup]
+        }
+        cur = corrected;                                                       // [n_embd, n_tokens, n_altup]
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;  // [n_embd, n_tokens, n_altup]
+
+    // cur now has multiple altup(s), we want to merge them back to 1 altup
+    {
+        ggml_tensor * target_magnitude = calc_magnitude(view_2d_slice(cur, i_altup_act));  // [n_embd, n_tokens]
+        // do a view to skip the first slice (active altup)
+        ggml_tensor * alt_slice =
+            ggml_view_3d(ctx0, cur, n_embd, n_tokens, n_altup - 1, ggml_row_size(cur->type, n_embd),
+                         ggml_row_size(cur->type, n_embd * n_tokens), n_embd * n_tokens * ggml_element_size(cur));
+        ggml_tensor * altup_unembd =
+            ggml_mul_mat(ctx0, model.altup_unembd_proj, alt_slice);  // shape: [n_embd, n_tokens, n_altup - 1]
+        ggml_tensor * new_magnitude = calc_magnitude(altup_unembd);
+        altup_unembd                = ggml_div(ctx0, ggml_mul(ctx0, altup_unembd, target_magnitude), new_magnitude);
+        cb(altup_unembd, "altup_unembd", -1);
+
+        // equivalent to torch.mean(hidden_states, dim=0)
+        cur = view_2d_slice(cur, 0);  // [n_embd, n_tokens]
+        for (int i = 0; i < n_altup - 1; ++i) {
+            cur = ggml_add(ctx0, cur, view_2d_slice(altup_unembd, i));
+        }
+        cur = ggml_scale(ctx0, cur, 1.0f / float(n_altup));  // [n_embd, n_tokens]
+        cb(cur, "unembd_merged", -1);
+    }
+    // cur now has shape: [n_embd, n_tokens]
+
+    // TODO: move this to right after the last KV layer
+    {
+        // skip computing output for unused tokens
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+        cur                       = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    {
+        // final logit soft-capping
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_gemma3n_iswa::calc_magnitude(ggml_tensor * x) {
+    return ggml_sqrt(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, x)));
+}
+
+// get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim
+ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
+    GGML_ASSERT(idx < (int) x->ne[2]);
+    return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1], ggml_row_size(x->type, x->ne[0]),
+                        idx * x->ne[0] * x->ne[1] * ggml_element_size(x));
+}
+
+// equivalent to get_per_layer_inputs() in python code
+// output shape: [n_embd_altup, n_layer, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
+    auto          inp = std::make_unique<llm_graph_input_embd>();
+    ggml_tensor * inp_per_layer;
+    if (ubatch.token) {
+        inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
+        ggml_set_input(inp->tokens);
+        res->t_tokens = inp->tokens;
+        inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
+        inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
+        inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup));
+        cb(inp_per_layer, "inp_per_layer_selected", -1);
+    } else {
+        GGML_ABORT("TODO: support embd input");
+    }
+    res->add_input(std::move(inp));
+    return inp_per_layer;
+}
+
+// equivalent to project_per_layer_inputs() in python code
+// this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim
+// output shape: [n_embd_altup, n_tokens, n_layer]
+ggml_tensor * llm_build_gemma3n_iswa::project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer) {
+    const float per_layer_projection_scale = 1.0f / sqrtf((float) n_embd);
+    const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);
+
+    ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds);
+    per_layer_proj               = ggml_scale(ctx0, per_layer_proj, per_layer_projection_scale);
+    per_layer_proj               = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_altup, n_layer, n_tokens);
+    per_layer_proj               = build_norm(per_layer_proj, model.per_layer_proj_norm, NULL, LLM_NORM_RMS,
+                                              -1);  // [n_embd_altup, n_layer, n_tokens]
+    cb(per_layer_proj, "per_layer_proj", -1);
+
+    inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj);
+    inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
+    cb(inp_per_layer, "inp_per_layer", -1);
+
+    // permute to shape: [n_embd_altup, n_tokens, n_layer]
+    inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3));
+    return inp_per_layer;
+}
+
+// input cur shape: [n_altup, n_tokens]
+// output    shape: [n_altup, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::laurel(ggml_tensor * cur, int il) {
+    ggml_tensor * tmp = cur;
+    tmp               = build_lora_mm(model.layers[il].laurel_l, tmp);
+    tmp               = build_lora_mm(model.layers[il].laurel_r, tmp);
+    tmp               = build_norm(tmp, model.layers[il].laurel_post_norm, NULL, LLM_NORM_RMS, il);
+    tmp               = ggml_add(ctx0, tmp, cur);
+    cb(tmp, "laurel_out", il);
+    return tmp;
+}
+
+// input x shape: [n_embd, n_tokens]
+// output  shape: [n_embd, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::gaussian_topk(ggml_tensor * x) {
+    ggml_tensor * mean = ggml_mean(ctx0, x);
+    ggml_tensor * std  = ggml_sqrt(ctx0, ggml_scale(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))),
+                                                    1.0f / (float) (x->ne[0] - 1)));
+    ggml_tensor * cutoff_x = ggml_add(ctx0, mean, ggml_scale(ctx0, std, f_sparsity_std_mul));
+    return ggml_relu(ctx0, ggml_sub(ctx0, x, cutoff_x));
+}
+
+//
+// altup functions
+//
+
+// equivalent to compute_router_modalities() in python code
+// input x shape: [n_embd,  n_tokens]
+// output  shape: [n_altup, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::altup_compute_router_modalities(ggml_tensor * x, int il) {
+    ggml_tensor * router_inputs = build_norm(x, model.layers[il].altup_router_norm, NULL, LLM_NORM_RMS, il);
+
+    // router_input_scale
+    router_inputs = ggml_scale(ctx0, router_inputs, 1.0f / (float) n_embd);
+
+    ggml_tensor * output = ggml_mul_mat(ctx0, model.layers[il].altup_router, router_inputs);
+    return ggml_tanh(ctx0, output);  // [n_altup, n_tokens]
+}
+
+// input cur shape: [n_embd, n_tokens, n_altup]
+// output    shape: [n_embd, n_tokens, n_altup]
+ggml_tensor * llm_build_gemma3n_iswa::altup_predict(ggml_tensor * cur, int il) {
+    ggml_tensor * activated  = view_2d_slice(cur, i_altup_act);                 // [n_embd, n_tokens]
+    ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
+    cb(modalities, "modalities", il);
+
+    ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_predict_coef, modalities);
+    cb(all_coefs, "all_coefs", il);
+    // first dim now having n_altup^2 elements, we reshape it to 2D (so we end up with 3D tensor)
+    all_coefs = ggml_reshape_3d(ctx0, all_coefs, n_altup, n_altup, n_tokens);
+
+    // permute to [n_altup, n_embd, n_tokens]
+    ggml_tensor * cur_permuted = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3));
+    ggml_tensor * predictions  = ggml_mul_mat(ctx0, cur_permuted, all_coefs);  // [n_altup, n_embd, n_tokens]
+
+    // final shape must be the same as cur: [n_embd, n_tokens, n_altup]
+    predictions = ggml_cont(ctx0, ggml_permute(ctx0, predictions, 0, 2, 1, 3));
+    predictions = ggml_add(ctx0, predictions, cur);
+    cb(predictions, "predictions", il);
+
+    return predictions;
+}
+
+// input predictions       shape: [n_embd, n_tokens, n_altup]
+// input activated         shape: [n_embd, n_tokens]
+// output                  shape: [n_embd, n_tokens, n_altup]
+ggml_tensor * llm_build_gemma3n_iswa::altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) {
+    ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
+    cb(modalities, "modalities", il);
+
+    ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);
+    ggml_tensor * innovation        = ggml_sub(ctx0, activated, active_prediction);  // [n_embd, n_tokens]
+    cb(innovation, "innovation", il);
+
+    ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities);  // [n_altup, n_tokens]
+    all_coefs               = ggml_scale_bias(ctx0, all_coefs, 1.0f, 1.0f);                    // + 1.0
+    cb(all_coefs, "all_coefs", il);
+    all_coefs = ggml_transpose(ctx0, all_coefs);                                               // [n_tokens, n_altup]
+    all_coefs = ggml_cont_3d(ctx0, all_coefs, 1, n_tokens, n_altup);                           // [1, n_tokens, n_altup]
+
+    innovation              = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1);
+    ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs);   // [n_embd, n_tokens, n_altup]
+    corrected               = ggml_add(ctx0, corrected, predictions);  // [n_embd, n_tokens, n_altup]
+    cb(corrected, "corrected", il);
+
+    return corrected;
+}

src/models/glm4-moe.cpp

@@ -0,0 +1,153 @@
+#include "models.h"
+
+llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // Only process up to last layer (skip final NextN layer)
+    // Final layer tensors are loaded but not processed in forward pass
+    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+    for (int il = 0; il < n_transformer_layers; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // Pre-attention norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            }
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            }
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            }
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            // Apply Q/K norm if available (GLM-4.5 355B variant)
+            if (model.layers[il].attn_q_norm) {
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_normed", il);
+            }
+            if (model.layers[il].attn_k_norm) {
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_normed", il);
+            }
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_transformer_layers - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // Post-attention norm
+        cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "post_attn_norm", il);
+
+        // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
+        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+            // Dense FFN layer
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // Process routed experts using existing MoE infrastructure
+            ggml_tensor * routed_out = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, hparams.expert_weights_norm,
+                    true, hparams.expert_weights_scale,
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il);
+            cb(routed_out, "ffn_moe_out", il);
+
+            // Process shared expert on original input
+            ggml_tensor * shared_out = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(shared_out, "ffn_shexp_out", il);
+
+            // Final output: routed_output + shared_output
+            cur = ggml_add(ctx0, routed_out, shared_out);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/glm4.cpp

@@ -0,0 +1,127 @@
+#include "models.h"
+
+
+
+llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // Pre-attention norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv == nullptr) {
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            } else {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
+                                    0 * sizeof(float) * (n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+            }
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // Post-attention norm (new!)
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "post_attn_norm", il);
+
+        // Add the input (residual connection after post-attention norm)
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            // Pre-MLP norm
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // MLP
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    NULL, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            // Post-MLP norm
+            cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "post_mlp_norm", il);
+        }
+        // Add residual connection after post-MLP norm
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        cb(inpL, "l_out", il);
+    }
+    // Final norm
+    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // Output projection
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gpt2.cpp

@@ -0,0 +1,105 @@
+#include "models.h"
+
+llm_build_gpt2::llm_build_gpt2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * pos;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
+    cb(pos, "pos_embd", -1);
+
+    inpL = ggml_add(ctx0, inpL, pos);
+    cb(inpL, "inpL", -1);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gptneox.cpp

@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+llm_build_gptneox::llm_build_gptneox(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // ffn
+        if (hparams.use_par_res) {
+            // attention and ffn are computed in parallel
+            // x = x + attn(ln1(x)) + ffn(ln2(x))
+
+            ggml_tensor * attn_out = cur;
+
+            cur = build_norm(inpL,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, inpL);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, attn_out);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        } else {
+            // attention and ffn are computed sequentially
+            // x = x + attn(ln1(x))
+            // x = x + ffn(ln2(x))
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/granite-hybrid.cpp

@@ -0,0 +1,196 @@
+#include "models.h"
+
+
+llm_build_granite_hybrid::llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp = build_inp_mem_hybrid();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // Positional embeddings populated if rope enabled
+    ggml_tensor * inp_pos = nullptr;
+    if (hparams.rope_finetuned) {
+        inp_pos = build_inp_pos();
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        if (hparams.is_recurrent(il)) {
+            // ssm layer //
+            cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
+        } else {
+            // attention layer //
+            cur = build_attention_layer(cur, inp_pos, inp->get_attn(), model, n_embd_head, il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        // ffn
+        cur = build_layer_ffn(cur, inpSA, model, il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    // For Granite architectures - scale logits
+    if (hparams.f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_granite_hybrid::build_attention_layer(ggml_tensor *             cur,
+                                                              ggml_tensor *             inp_pos,
+                                                              llm_graph_input_attn_kv * inp_attn,
+                                                              const llama_model &       model,
+                                                              const int64_t             n_embd_head,
+                                                              const int                 il) {
+    // compute Q and K and (optionally) RoPE them
+    ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+    cb(Qcur, "Qcur", il);
+    if (model.layers[il].bq) {
+        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+        cb(Qcur, "Qcur", il);
+    }
+
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    cb(Kcur, "Kcur", il);
+    if (model.layers[il].bk) {
+        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+        cb(Kcur, "Kcur", il);
+    }
+
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    cb(Vcur, "Vcur", il);
+    if (model.layers[il].bv) {
+        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+        cb(Vcur, "Vcur", il);
+    }
+
+    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens);
+    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+    const bool use_rope = hparams.rope_finetuned;
+    if (use_rope) {
+        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+    }
+
+    cb(Qcur, "Qcur", il);
+    cb(Kcur, "Kcur", il);
+    cb(Vcur, "Vcur", il);
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+    cur = build_attn(inp_attn,
+            model.layers[il].wo, model.layers[il].bo,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+    cb(cur, "attn_out", il);
+    return cur;
+}
+
+ggml_tensor * llm_build_granite_hybrid::build_layer_ffn(ggml_tensor *       cur,
+                                                        ggml_tensor *       inpSA,
+                                                        const llama_model & model,
+                                                        const int           il) {
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "ffn_inp", il);
+
+    // feed-forward network (non-MoE)
+    if (model.layers[il].ffn_gate_inp == nullptr) {
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+    } else {
+        // MoE branch
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // For Granite MoE Shared
+        if (hparams.n_ff_shexp > 0) {
+            ggml_tensor * ffn_shexp =
+                build_ffn(cur,
+                    model.layers[il].ffn_up_shexp, NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = moe_out;
+        }
+    }
+
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "ffn_out", il);
+
+    cur = build_cvec(cur, il);
+    cb(cur, "l_out", il);
+
+    return cur;
+}