sycl : enhance the api to support peer-to-peer copy (#27550)

* convert : skip bias_vl tensor in DeepSeek-V4 DSpark conversion

The DFLASH arch does not include FFN_EXP_PROBS_B_VL, so the DSpark
conversion failed when it tried to write the mtmd-only hash routing
tensor ffn.gate.bias_vl. Drop it like the tid2eid tensor; the DFLASH
draft only consumes ffn.gate.bias via FFN_EXP_PROBS_B.

Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731

* cont : fix

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

.github/actions/windows-setup-rocm/action.yml

@@ -24,7 +24,7 @@ runs:
 
         write-host "Installing ROCm wheels for multi-arch support"
         # Install ROCm wheels for multi-arch support (this may take several minutes)
-        python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ inputs.version }}"
+        python -m pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel]==${{ inputs.version }}"
 
         # Pre-expand the devel tree so it is included in the cache
         write-host "Initializing ROCm devel tree"

.github/workflows/build-apple.yml

@@ -66,7 +66,13 @@ jobs:
             -DGGML_RPC=ON \
             -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
           time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
-          leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1
+
+      - name: Check for leaks
+        run: |
+          cmd=(./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1)
+          leaks -atExit -- "${cmd[@]}"
+          # Graphics devices are leaked by Metal in Apple code sometimes, so we ignore those leaks
+          OBJC_DEBUG_MISSING_POOLS=YES "${cmd[@]}" 2>&1 | awk '{ print } index($0, "autoreleased with no pool in place") && !/class [a-zA-Z0-9]+Device autoreleased/ { found = 1 } END { exit found }'
 
       - name: Test
         id: cmake_test

.github/workflows/release.yml

@@ -725,7 +725,7 @@ jobs:
     strategy:
       matrix:
         include:
-          - ROCM_VERSION: "7.14.0"
+          - ROCM_VERSION: "10.0.0"
             gpu_targets: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
             build: x64
 
@@ -1279,7 +1279,7 @@ jobs:
     strategy:
       matrix:
         include:
-          - ROCM_VERSION: "7.14.0"
+          - ROCM_VERSION: "10.0.0"
             gpu_targets: "gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
             build: 'x64'
 
@@ -1333,7 +1333,7 @@ jobs:
           # libraries = HIP runtime and CMake configs needed for linking
           # devel = compilers, headers, static libs
           python -m pip install --upgrade pip
-          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"
+          python -m pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"
 
           # Get ROCm installation paths using the rocm-sdk CLI tool
           ROCM_PATH=$(rocm-sdk path --root)
@@ -1703,7 +1703,7 @@ jobs:
             - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
             - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
             - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
-            - [Ubuntu x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.14-x64.tar.gz)
+            - [Ubuntu x64 (ROCm 10.0)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-10.0-x64.tar.gz)
             - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
             - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
             - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
@@ -1721,7 +1721,7 @@ jobs:
             - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)
             - [Windows x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ needs.windows-openvino.outputs.openvino_version }}-x64.zip)
             - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip)
-            - [Windows x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-rocm-7.14-x64.zip)
+            - [Windows x64 (ROCm 10.0)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-rocm-10.0-x64.zip)
 
             **openEuler:**
             - [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23705)

.github/workflows/server-self-hosted.yml

@@ -102,7 +102,7 @@ jobs:
           ./tests.sh
 
   server-cuda:
-    runs-on: [self-hosted, llama-server, Linux, NVIDIA]
+    runs-on: "hf-jobs-t4-small:cuda13"
 
     steps:
       - name: Clone
@@ -112,12 +112,42 @@ jobs:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y cmake libssl-dev python3 python3-venv python3-pip
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.24
+        with:
+          restore: false
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        with:
+          key: self-hosted-server-cuda
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
       - name: Build
         id: cmake_build
         run: |
-          cmake -B build -DGGML_CUDA=ON -DGGML_SCHED_NO_REALLOC=ON
+          cmake -B build -DGGML_CUDA=ON -DGGML_SCHED_NO_REALLOC=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc
           cmake --build build --config Release -j $(nproc) --target llama-server
 
+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: self-hosted-server-cuda
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
       - name: Python setup
         id: setup_python
         run: |

common/arg.cpp

@@ -960,6 +960,11 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         ));
     }
 
+    // if the preserve_reasoning kwarg was not specified explicitly, enable it by default
+    if (!params.default_template_kwargs.count("preserve_reasoning")) {
+        params.default_template_kwargs["preserve_reasoning"] = "true";
+    }
+
     return true;
 }
 
@@ -3553,6 +3558,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                     LOG_WRN("Setting 'enable_thinking' via --chat-template-kwargs is deprecated. "
                             "Use --reasoning on / --reasoning off instead.\n");
                 }
+                if (item.key() == "preserve_reasoning") {
+                    LOG_WRN("Setting 'preserve_reasoning' via --chat-template-kwargs is deprecated. "
+                            "Use --reasoning-preserve / --no-reasoning-preserve instead.\n");
+                }
                 params.default_template_kwargs[item.key()] = item.value().dump();
             }
         }
@@ -3743,7 +3752,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"--reasoning-preserve"},
         {"--no-reasoning-preserve"},
-        "preserve reasoning trace in the full history, not just the last assistant message (default: template default)\n"
+        "preserve reasoning trace in the full history, not just the last assistant message (default: enabled)\n"
         "compatible with certain templates having 'supports_preserve_reasoning' capability\n"
         "example: https://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking",
         [](common_params & params, bool value) {
@@ -3752,6 +3761,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             } else {
                 params.default_template_kwargs["preserve_reasoning"] = "false";
             }
+            params.preserve_reasoning_specified = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_REASONING_PRESERVE"));
     add_opt(common_arg(

common/common.h

@@ -270,7 +270,7 @@ struct common_params_sampling {
         COMMON_SAMPLER_TYPE_TEMPERATURE,
     };
 
-    common_grammar              grammar;      // optional grammar constraint (user / output-format / tool-calls)
+    common_grammar                      grammar;          // optional grammar constraint (user / output-format / tool-calls)
     bool                                grammar_lazy = false;
     std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
     std::set<llama_token>               preserved_tokens;
@@ -657,6 +657,7 @@ struct common_params {
     std::string ssl_file_cert = "";                                                                         // NOLINT
 
     std::map<std::string, std::string> default_template_kwargs;
+    bool preserve_reasoning_specified = false;
 
     // CLI params
     std::string server_base; // if set, connect to this server instead of starting a new one

conversion/__init__.py

@@ -188,6 +188,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "NanbeigeForCausalLM": "nanbeige",
     "NemotronForCausalLM": "nemotron",
     "NemotronHForCausalLM": "nemotron",
+    "NemotronHPuzzleForCausalLM": "nemotron",
     "NeoBERT": "bert",
     "NeoBERTForSequenceClassification": "bert",
     "NeoBERTLMHead": "bert",

conversion/deepseek.py

@@ -578,8 +578,7 @@ class DeepseekV4Model(TextModel):
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, gen = item
-        if (name.startswith(("aligner.", "image_"))
-                or name.endswith(".ffn.gate.bias_vl")):
+        if name.startswith(("aligner.", "image_")):
             return None
         if name.startswith("mtp."):
             if not cls.mtp_only:
@@ -856,6 +855,7 @@ class DeepseekV4Model(TextModel):
             "ffn_norm.weight": (gguf.MODEL_TENSOR.FFN_NORM, ".weight"),
             "ffn.gate.weight": (gguf.MODEL_TENSOR.FFN_GATE_INP, ".weight"),
             "ffn.gate.bias": (gguf.MODEL_TENSOR.FFN_EXP_PROBS_B, ".bias"),
+            "ffn.gate.bias_vl": (gguf.MODEL_TENSOR.FFN_EXP_PROBS_B_VL, ".bias"),
             "ffn.gate.tid2eid": (gguf.MODEL_TENSOR.FFN_GATE_TID2EID, ".weight"),
             "ffn.shared_experts.w1.weight": (gguf.MODEL_TENSOR.FFN_GATE_SHEXP, ".weight"),
             "ffn.shared_experts.w2.weight": (gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, ".weight"),
@@ -881,6 +881,10 @@ class DeepseekV4Model(TextModel):
         if re.match(r"layers\.\d+\.ffn\.experts\.\d+\.w[123]\.(weight|scale)$", name):
             return []
 
+        # hash layers route text tokens via tid2eid and image tokens via bias_vl; gate.bias is unused
+        if name.endswith(".ffn.gate.bias") and bid is not None and bid < self.hparams["num_hash_layers"]:
+            return []
+
         tensor_key, suffix = self._map_dsv4_tensor_name(name, bid)
         if tensor_key == gguf.MODEL_TENSOR.FFN_GATE_TID2EID:
             return []
@@ -1003,6 +1007,13 @@ class DeepseekV4DSparkModel(DeepseekV4Model):
             return self._DSPARK_ROOT_MAP[name]
         return super()._map_dsv4_tensor_name(name, bid)
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # the DFlash draft uses the plain exp-probs bias (ffn.gate.bias -> FFN_EXP_PROBS_B);
+        # the mtmd-only hash routing tensors (bias_vl, tid2eid) are not part of the DFLASH arch
+        if name.endswith(".ffn.gate.bias_vl"):
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
+
     def set_vocab(self):
         if self.target_model_dir is None:
             raise ValueError("DeepSeek-V4 DSpark requires --target-model-dir with the target tokenizer")

conversion/nemotron.py

@@ -5,6 +5,7 @@ from typing import Any, Callable, Iterable, TYPE_CHECKING
 import torch
 
 if TYPE_CHECKING:
+    from pathlib import Path
     from torch import Tensor
 
 from .base import MmprojModel, ModelBase, TextModel, gguf, logger
@@ -201,6 +202,7 @@ class NemotronHModel(GraniteHybridModel):
     model_arch = gguf.MODEL_ARCH.NEMOTRON_H
     is_moe: bool = False
     supports_mtp_export = True
+    _experts: list[dict[str, Tensor]] | None = None
 
     _SSM_LAYER_TYPES = {"mamba", "linear_attention"}
     _ATTN_LAYER_TYPES = {"attention", "full_attention"}
@@ -513,3 +515,88 @@ class NemotronHModel(GraniteHybridModel):
             experts = [k for d in self._experts for k in d.keys()]
             if len(experts) > 0:
                 raise ValueError(f"Unprocessed experts: {experts}")
+
+
+@ModelBase.register("NemotronHPuzzleForCausalLM")
+@ModelBase.example("nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16")
+class NemotronHPuzzleModel(NemotronHModel):
+    """NVIDIA Puzzle: NemotronH with a per-block MoE config (block_configs).
+
+    The checkpoint also ships an MTP draft head (mtp.safetensors). It is skipped
+    here: there is no Puzzle MTP inference path in tree, and the head is laid out
+    by mtp_block_configs rather than the mtp.layers.* form NemotronHModel maps."""
+
+    model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
+    is_moe: bool = True
+    supports_mtp_export = False
+
+    def __init__(self, dir_model: "Path", *args, **kwargs):
+        hparams = dict(kwargs.pop("hparams", None) or ModelBase.load_hparams(dir_model, self.is_mistral_format))
+
+        self.block_configs: list[dict] = hparams["block_configs"]
+        self.n_layer_trunk = len(self.block_configs)
+
+        # block_configs carries the per-block MoE shape, and is the authority on the
+        # block pattern too: the layers_block_type the HF config wrapper computes is
+        # not sized to it.
+        hparams["num_hidden_layers"] = self.n_layer_trunk
+        hparams["layers_block_type"] = [bc["block_type"] for bc in self.block_configs]
+
+        self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
+
+        # Bypass NemotronHModel.__init__: it assumes a flat num_experts_per_tok /
+        # moe_intermediate_size and a layers_block_type sized to block_count, neither
+        # of which hold for Puzzle's per-block config.
+        GraniteHybridModel.__init__(self, dir_model, *args, hparams=hparams, **kwargs)
+
+        self.head_dim = self.find_hparam(["head_dim", "attention_head_dim"])
+        self.d_inner = self.find_hparam(["num_heads"]) * self.d_model
+
+        # NemotronHModel.__init__ folds an MTP block into block_count when the
+        # config carries num_nextn_predict_layers; Puzzle's config does, but its
+        # head has a different layout and no inference path, so stay opted out.
+        self._mtp_bid = None
+
+    def set_gguf_parameters(self):
+        GraniteHybridModel.set_gguf_parameters(self)
+
+        head_dim = self.head_dim
+        if head_dim is None:
+            raise ValueError("Could not find the attention head dim in config")
+        self.gguf_writer.add_key_length(head_dim)
+        self.gguf_writer.add_value_length(head_dim)
+
+        ffn_lengths = [bc.get("moe_intermediate_size") or 0 for bc in self.block_configs]
+        experts_used = [bc.get("num_experts_per_tok") or 0 for bc in self.block_configs]
+
+        self.gguf_writer.add_feed_forward_length(ffn_lengths)
+        self.gguf_writer.add_expert_feed_forward_length(ffn_lengths)
+        self.gguf_writer.add_expert_used_count(experts_used)
+
+        self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["moe_shared_expert_intermediate_size"])
+        self.gguf_writer.add_expert_count(self.hparams["n_routed_experts"])
+        self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
+        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_group_count(self.hparams["n_group"])
+        self.gguf_writer.add_moe_latent_size(self.hparams["moe_latent_size"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # The official BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16)
+        # names the trunk "model.*" (model.layers.*, model.embeddings, model.norm_f)
+        # where the original release used the NemotronH-style "backbone.*", and spells
+        # the router bias "e_score_correction_bias" instead of "e_score_correction.bias";
+        # normalize so both convert identically.
+        if name.startswith("model."):
+            name = "backbone." + name[len("model."):]
+        if name.endswith("mixer.gate.e_score_correction_bias"):
+            name = name[: -len("e_score_correction_bias")] + "e_score_correction.bias"
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        # Drop the MTP head unconditionally; see the class docstring.
+        if item[0].startswith("mtp."):
+            return None
+        return super().filter_tensors(item)

examples/training/README.md

@@ -6,6 +6,8 @@ Finetuning of Stories 260K and LLaMA 3.2 1b seems to work with 24 GB of memory.
 **For CPU training, compile llama.cpp without any additional backends such as CUDA.**
 **For CUDA training, use the maximum number of GPU layers.**
 
+Flash attention is disabled during training because `FLASH_ATTN_EXT` has no backward pass.
+
 Proof of concept:
 
 ``` sh

ggml/src/ggml-cuda/mmq-vec-dot.cuh

@@ -148,7 +148,6 @@ static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma(
     typedef tile<16,  8, int, input_layout>        tile_B;
     typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -204,7 +203,6 @@ static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma(
     typedef tile< 8, 8, int> tile_B;
     typedef tile<16, 8, int> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -320,7 +318,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
     typedef tile<16,  8, int, input_layout>        tile_B;
     typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -371,7 +368,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
     typedef tile< 8,  8, int> tile_B;
     typedef tile<16,  8, int> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -486,7 +482,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
     typedef tile<16,  4, int, input_layout>        tile_B;
     typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -537,7 +532,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
     typedef tile< 8, 4, int> tile_B;
     typedef tile<16, 8, int> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -686,7 +680,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
     typedef tile<16,  4, int, input_layout>        tile_B;
     typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -756,7 +749,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
     typedef tile< 8, 4, int> tile_B;
     typedef tile<16, 8, int> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -1023,7 +1015,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
     typedef tile<16,  4, int, input_layout>        tile_B;
     typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -1075,7 +1066,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
     typedef tile< 8, 4, int> tile_B;
     typedef tile<16, 8, int> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -1190,7 +1180,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
     typedef tile<8,  8, int>   tile_B;
     typedef tile<16, 8, float> tile_C;
 
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp / tile_C::I;

ggml/src/ggml-cuda/mmq.cuh

@@ -481,9 +481,6 @@ static __device__ __forceinline__ void ggml_cuda_mmq_write_back_mma(
     typedef tile<16,  8, int> tile_C;
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
-    constexpr int warp_size     = ggml_cuda_get_physical_warp_size();
-    constexpr int nwarps        = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
     constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
     constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
 
@@ -540,8 +537,6 @@ struct ggml_cuda_mmq_util_funcs {
 
 template <ggml_type type, int J, bool fallback>
 static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_funcs() {
-    constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
-
     if (!ggml_cuda_mmq_get_config(type, J, fallback).use_mma_data_layout()) {
         switch (type) {
             case GGML_TYPE_Q1_0:

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -4005,8 +4005,10 @@ static void ggml_hexagon_precompute_unary_params(
 
     kparams->n_threads = n_threads;
 
-    const size_t src0_data_row_size = src0->ne[0] * sizeof(float);
-    const size_t dst_data_row_size  = dst->ne[0]  * sizeof(float);
+    const size_t elem_size = ggml_type_size(src0->type);
+
+    const size_t src0_data_row_size = src0->ne[0] * elem_size;
+    const size_t dst_data_row_size  = dst->ne[0]  * ggml_type_size(dst->type);
 
     const size_t src0_row_size_aligned = hex_round_up(src0_data_row_size, 128);
     const size_t dst_row_size_aligned  = hex_round_up(dst_data_row_size,  128);
@@ -4020,7 +4022,7 @@ static void ggml_hexagon_precompute_unary_params(
 
     if (op == HTP_OP_RMS_NORM_MUL) {
         GGML_ASSERT(src1 != nullptr);
-        src1_data_row_size = src1->ne[0] * sizeof(float);
+        src1_data_row_size = src1->ne[0] * ggml_type_size(src1->type);
         src1_row_size_aligned = hex_round_up(src1_data_row_size, 128);
         broadcast_weight = (src1->ne[1] * src1->ne[2] * src1->ne[3] == 1);
     }
@@ -4034,7 +4036,7 @@ static void ggml_hexagon_precompute_unary_params(
 
     htp_unary_vtcm_layout_build(&L, op, src0->ne[0], dst->ne[0],
                                 op == HTP_OP_RMS_NORM_MUL ? src1->ne[0] : 0,
-                                broadcast_weight, n_threads, sess->vtcm_size,
+                                broadcast_weight, n_threads, sess->vtcm_size, elem_size,
                                 &col_tile, &vtcm_row_per_thread);
 
     kparams->col_tile = col_tile;
@@ -4451,15 +4453,39 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
     const struct ggml_tensor * src0 = op->src[0];
     const struct ggml_tensor * dst  = op;
 
-    if (src0->type != GGML_TYPE_F32) {
+    if (src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) {
         return false;
     }
-    if (dst->type != GGML_TYPE_F32) {
+    if (dst->type != src0->type) {
         return false;
     }
     if (!ggml_is_contiguous_rows(src0)) {
         return false;
     }
+
+    // F16 device kernels only cover this explicit whitelist (must stay in sync with
+    // the is_f16 whitelist in execute_op_unary(), unary-ops.c).
+    if (src0->type == GGML_TYPE_F16) {
+        switch (op->op) {
+            case GGML_OP_NORM:
+            case GGML_OP_RMS_NORM:
+            case GGML_OP_L2_NORM:
+            case GGML_OP_SCALE:
+            case GGML_OP_CLAMP:
+            case GGML_OP_SQR:
+            case GGML_OP_SQRT:
+            case GGML_OP_LOG:
+                break;
+            case GGML_OP_UNARY:
+                if (ggml_get_unary_op(op) != GGML_UNARY_OP_ABS) {
+                    return false;
+                }
+                break;
+            default:
+                return false;
+        }
+    }
+
     if (!ggml_are_same_shape(src0, dst)) {
         return false;
     }

ggml/src/ggml-hexagon/htp/hvx-arith.h

@@ -358,6 +358,54 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
     }
 }
 
+#define HVX_OP_CLAMP_SCALAR_F16(v)                                       \
+    ({                                                                   \
+        HVX_VectorPred pred_cap_right = Q6_Q_vcmp_gt_VhfVhf(v, max_vec); \
+        HVX_VectorPred pred_cap_left  = Q6_Q_vcmp_gt_VhfVhf(min_vec, v); \
+        HVX_Vector tmp = Q6_V_vmux_QVV(pred_cap_right, max_vec, v);      \
+        Q6_V_vmux_QVV(pred_cap_left, min_vec, tmp);                      \
+    })
+
+static inline void hvx_clamp_scalar_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 min, const _Float16 max, uint32_t n) {
+    const HVX_Vector min_vec = hvx_vec_splat_f16(min);
+    const HVX_Vector max_vec = hvx_vec_splat_f16(max);
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(_Float16), hvx_vec_store_a, HVX_OP_CLAMP_SCALAR_F16);
+}
+
+static inline void hvx_clamp_scalar_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 min, const _Float16 max, uint32_t n) {
+    const HVX_Vector min_vec = hvx_vec_splat_f16(min);
+    const HVX_Vector max_vec = hvx_vec_splat_f16(max);
+    assert((unsigned long) dst % 128 == 0);
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(_Float16), hvx_vec_store_a, HVX_OP_CLAMP_SCALAR_F16);
+}
+
+static inline void hvx_clamp_scalar_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 min, const _Float16 max, uint32_t n) {
+    const HVX_Vector min_vec = hvx_vec_splat_f16(min);
+    const HVX_Vector max_vec = hvx_vec_splat_f16(max);
+    assert((unsigned long) src % 128 == 0);
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(_Float16), hvx_vec_store_u, HVX_OP_CLAMP_SCALAR_F16);
+}
+
+static inline void hvx_clamp_scalar_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 min, const _Float16 max, uint32_t n) {
+    const HVX_Vector min_vec = hvx_vec_splat_f16(min);
+    const HVX_Vector max_vec = hvx_vec_splat_f16(max);
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(_Float16), hvx_vec_store_u, HVX_OP_CLAMP_SCALAR_F16);
+}
+
+static inline void hvx_clamp_scalar_f16(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 min, const _Float16 max, const int num_elems) {
+    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {
+        hvx_clamp_scalar_f16_aa(dst, src, min, max, num_elems);
+    } else if (hex_is_aligned((void *) dst, 128)) {
+        hvx_clamp_scalar_f16_au(dst, src, min, max, num_elems);
+    } else if (hex_is_aligned((void *) src, 128)) {
+        hvx_clamp_scalar_f16_ua(dst, src, min, max, num_elems);
+    } else {
+        hvx_clamp_scalar_f16_uu(dst, src, min, max, num_elems);
+    }
+}
+
 //
 // Abs
 //
@@ -386,11 +434,69 @@ static inline void hvx_abs_f32_aa(uint8_t * restrict dst, const uint8_t * restri
     }
 }
 
+#define hvx_abs_f16_loop_body(dst_type, src_type, vec_store)               \
+    do {                                                                   \
+        dst_type * restrict vdst  = (dst_type *) dst;                      \
+        src_type * restrict vsrc = (src_type *) src;                       \
+                                                                           \
+        const uint32_t elem_size = sizeof(_Float16);                       \
+        const uint32_t epv  = 128 / elem_size;                             \
+        const uint32_t nvec = n / epv;                                     \
+        const uint32_t nloe = n % epv;                                     \
+                                                                           \
+        uint32_t i = 0;                                                    \
+                                                                           \
+        _Pragma("unroll(4)")                                               \
+        for (; i < nvec; i++) {                                            \
+            vdst[i] = hvx_vec_abs_f16(vsrc[i]);                            \
+        }                                                                  \
+        if (nloe) {                                                        \
+            HVX_Vector v = hvx_vec_abs_f16(vsrc[i]);                       \
+            vec_store((void *) &vdst[i], nloe * elem_size, v);             \
+        }                                                                  \
+    } while(0)
+
+static inline void hvx_abs_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_abs_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_abs_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_abs_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_abs_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_abs_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_abs_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_abs_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_abs_f16(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) {
+    if (hex_is_aligned((void *) dst, 128)) {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_abs_f16_aa(dst, src, num_elems);
+        } else {
+            hvx_abs_f16_au(dst, src, num_elems);
+        }
+    } else {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_abs_f16_ua(dst, src, num_elems);
+        } else {
+            hvx_abs_f16_uu(dst, src, num_elems);
+        }
+    }
+}
+
 //
 // Square
 //
 
-#define hvx_sqr_f32_loop_body(dst_type, src_type, vec_store)           \
+#define hvx_sqr_f32_loop_body(dst_type, src_type, vec_store)               \
     do {                                                                   \
         dst_type * restrict vdst  = (dst_type *) dst;                      \
         src_type * restrict vsrc = (src_type *) src;                       \
@@ -404,10 +510,10 @@ static inline void hvx_abs_f32_aa(uint8_t * restrict dst, const uint8_t * restri
                                                                            \
         _Pragma("unroll(4)")                                               \
         for (; i < nvec; i++) {                                            \
-            vdst[i] = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);                        \
+            vdst[i] = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);                    \
         }                                                                  \
         if (nloe) {                                                        \
-            HVX_Vector v = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);                   \
+            HVX_Vector v = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);               \
             vec_store((void *) &vdst[i], nloe * elem_size, v);             \
         }                                                                  \
     } while(0)
@@ -448,6 +554,64 @@ static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict
     }
 }
 
+#define hvx_sqr_f16_loop_body(dst_type, src_type, vec_store)               \
+    do {                                                                   \
+        dst_type * restrict vdst  = (dst_type *) dst;                      \
+        src_type * restrict vsrc = (src_type *) src;                       \
+                                                                           \
+        const uint32_t elem_size = sizeof(_Float16);                       \
+        const uint32_t epv  = 128 / elem_size;                             \
+        const uint32_t nvec = n / epv;                                     \
+        const uint32_t nloe = n % epv;                                     \
+                                                                           \
+        uint32_t i = 0;                                                    \
+                                                                           \
+        _Pragma("unroll(4)")                                               \
+        for (; i < nvec; i++) {                                            \
+            vdst[i] = HVX_OP_MUL_F16(vsrc[i], vsrc[i]);                    \
+        }                                                                  \
+        if (nloe) {                                                        \
+            HVX_Vector v = HVX_OP_MUL_F16(vsrc[i], vsrc[i]);               \
+            vec_store((void *) &vdst[i], nloe * elem_size, v);             \
+        }                                                                  \
+    } while(0)
+
+static inline void hvx_sqr_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqr_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqr_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_sqr_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqr_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqr_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqr_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_sqr_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqr_f16(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) {
+    if (hex_is_aligned((void *) dst, 128)) {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_sqr_f16_aa(dst, src, num_elems);
+        } else {
+            hvx_sqr_f16_au(dst, src, num_elems);
+        }
+    } else {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_sqr_f16_ua(dst, src, num_elems);
+        } else {
+            hvx_sqr_f16_uu(dst, src, num_elems);
+        }
+    }
+}
+
 #undef HVX_OP_ADD_F32
 #undef HVX_OP_SUB_F32
 #undef HVX_OP_MUL_F32
@@ -464,6 +628,7 @@ static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict
 #undef hvx_scalar_loop_body
 #undef HVX_OP_MIN_SCALAR
 #undef HVX_OP_CLAMP_SCALAR
+#undef HVX_OP_CLAMP_SCALAR_F16
 #undef DEFINE_HVX_BINARY_OP_VARIANTS
 #undef HVX_BINARY_DISPATCHER
 #undef UNUSED

ggml/src/ggml-hexagon/htp/hvx-log.h

@@ -86,4 +86,33 @@ static inline void hvx_log_f32_aa(uint8_t * restrict dst, const uint8_t * restri
     }
 }
 
+// Compute log(x) for f16 by promoting to f32, applying hvx_vec_log_f32, and narrowing back.
+static inline void hvx_log_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+
+    HVX_Vector * restrict vdst = (HVX_Vector *) dst;
+    HVX_Vector * restrict vsrc = (HVX_Vector *) src;
+
+    const uint32_t nvec = n / VLEN_FP16;
+    const uint32_t nloe = n % VLEN_FP16;
+
+    uint32_t i = 0;
+
+    _Pragma("unroll(4)")
+    for (; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);
+        HVX_Vector r0 = hvx_vec_log_f32(Q6_V_lo_W(p));
+        HVX_Vector r1 = hvx_vec_log_f32(Q6_V_hi_W(p));
+        vdst[i] = hvx_vec_f32_to_f16(r0, r1);
+    }
+    if (nloe) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);
+        HVX_Vector r0 = hvx_vec_log_f32(Q6_V_lo_W(p));
+        HVX_Vector r1 = hvx_vec_log_f32(Q6_V_hi_W(p));
+        HVX_Vector v = hvx_vec_f32_to_f16(r0, r1);
+        hvx_vec_store_a((void *) &vdst[i], nloe * SIZEOF_FP16, v);
+    }
+}
+
 #endif /* HVX_LOG_H */

ggml/src/ggml-hexagon/htp/hvx-norm.h

@@ -254,4 +254,201 @@ static inline void hvx_fast_l2_norm_f32(const uint8_t * restrict src,
     }
 }
 
+// F16 norm kernels: reduce and scale in f32 (via promote/narrow), matching the
+// precision-preserving pattern used by the flash-attn f16 kernels.
+
+static inline void hvx_fast_rms_norm_f16(const uint8_t * restrict src,
+                                         uint8_t * restrict dst,
+                                         const int num_elems,
+                                         float     epsilon) {
+
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    const int nvec = num_elems / VLEN_FP16;    // number of full f16 vectors
+    const int nloe = num_elems % VLEN_FP16;    // leftover elements
+
+    HVX_Vector sum_v = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+    }
+
+    sum_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
+
+    HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
+    HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
+    HVX_Vector mean_v         = Q6_Vqf32_vmpy_VsfVsf(sum_v, denom_v);
+    HVX_Vector mean_epsilon_v = Q6_Vqf32_vadd_Vqf32Vsf(mean_v, epsilon_v);
+
+    HVX_Vector scale_v = hvx_vec_rsqrt_f32(Q6_Vsf_equals_Vqf32(mean_epsilon_v));
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), scale_v));
+        v_dst[i] = hvx_vec_f32_to_f16(r0, r1);
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), scale_v));
+        HVX_Vector result = hvx_vec_f32_to_f16(r0, r1);
+        hvx_vec_store_a(&v_dst[nvec], nloe * SIZEOF_FP16, result);
+    }
+}
+
+static inline void hvx_fast_norm_f16(const uint8_t * restrict src,
+                                     uint8_t * restrict dst,
+                                     const int num_elems,
+                                     float     epsilon) {
+
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    const int nvec = num_elems / VLEN_FP16;
+    const int nloe = num_elems % VLEN_FP16;
+
+    HVX_Vector sum_sq_v = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector sum_x_v  = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_sq_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_sq_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_sq_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_sq_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+        sum_x_v  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_x_v,  Q6_Vqf32_vadd_VsfVsf(p0, Q6_V_vzero()));
+        sum_x_v  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_x_v,  Q6_Vqf32_vadd_VsfVsf(p1, Q6_V_vzero()));
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_sq_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_sq_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_sq_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_sq_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+        sum_x_v  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_x_v,  Q6_Vqf32_vadd_VsfVsf(p0, Q6_V_vzero()));
+        sum_x_v  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_x_v,  Q6_Vqf32_vadd_VsfVsf(p1, Q6_V_vzero()));
+    }
+
+    sum_sq_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_sq_v));
+    sum_x_v  = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_x_v));
+
+    HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
+    HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
+    HVX_Vector mean_sq_v      = Q6_Vqf32_vmpy_VsfVsf(sum_sq_v, denom_v);
+    HVX_Vector mean_x_v       = Q6_Vqf32_vmpy_VsfVsf(sum_x_v,  denom_v);
+    HVX_Vector mean_x_sq_v    = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(mean_x_v), Q6_Vsf_equals_Vqf32(mean_x_v));
+    HVX_Vector var_v          = Q6_Vqf32_vsub_Vqf32Vqf32(mean_sq_v, mean_x_sq_v);
+    HVX_Vector var_epsilon_v  = Q6_Vqf32_vadd_Vqf32Vsf(var_v, epsilon_v);
+
+    HVX_Vector scale_v  = hvx_vec_rsqrt_f32(Q6_Vsf_equals_Vqf32(var_epsilon_v));
+    HVX_Vector mean_x_b = hvx_vec_repl_f32(Q6_Vsf_equals_Vqf32(mean_x_v));
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector d0 = Q6_Vqf32_vsub_VsfVsf(Q6_V_lo_W(p), mean_x_b);
+        HVX_Vector d1 = Q6_Vqf32_vsub_VsfVsf(Q6_V_hi_W(p), mean_x_b);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(d0), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(d1), scale_v));
+        v_dst[i] = hvx_vec_f32_to_f16(r0, r1);
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector d0 = Q6_Vqf32_vsub_VsfVsf(Q6_V_lo_W(p), mean_x_b);
+        HVX_Vector d1 = Q6_Vqf32_vsub_VsfVsf(Q6_V_hi_W(p), mean_x_b);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(d0), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(d1), scale_v));
+        HVX_Vector result = hvx_vec_f32_to_f16(r0, r1);
+        hvx_vec_store_a(&v_dst[nvec], nloe * SIZEOF_FP16, result);
+    }
+}
+
+static inline void hvx_fast_l2_norm_f16(const uint8_t * restrict src,
+                                        uint8_t * restrict dst,
+                                        const int num_elems,
+                                        float     epsilon) {
+
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    const int nvec = num_elems / VLEN_FP16;
+    const int nloe = num_elems % VLEN_FP16;
+
+    HVX_Vector sum_v = hvx_vec_splat_f32(0.0f);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+    }
+
+    HVX_Vector sum_sf    = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
+    HVX_Vector rsqrt_v   = hvx_vec_rsqrt_f32(sum_sf);
+    HVX_Vector sqrt_v    = hvx_vec_inverse_f32(rsqrt_v);
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+    HVX_Vector denom_v   = Q6_Vsf_vmax_VsfVsf(sqrt_v, epsilon_v);
+    HVX_Vector scale_v   = hvx_vec_inverse_f32(denom_v);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), scale_v));
+        v_dst[i] = hvx_vec_f32_to_f16(r0, r1);
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), scale_v));
+        HVX_Vector result = hvx_vec_f32_to_f16(r0, r1);
+        hvx_vec_store_a(&v_dst[nvec], nloe * SIZEOF_FP16, result);
+    }
+}
+
 #endif // HVX_NORM_H

ggml/src/ggml-hexagon/htp/hvx-scale.h

@@ -130,4 +130,70 @@ static inline void hvx_scale_offset_f32(uint8_t * restrict dst, const uint8_t *
     }
 }
 
+// Scale+offset computed by promoting f16 -> f32, then narrowing the result back to f16.
+#define hvx_scale_offset_f16_loop_body(dst_type, src_type, vec_store)                                                \
+    do {                                                                                                             \
+        dst_type * restrict vdst = (dst_type *) dst;                                                                 \
+        src_type * restrict vsrc = (src_type *) src;                                                                 \
+                                                                                                                     \
+        HVX_Vector vs = hvx_vec_splat_f32(scale);                                                                    \
+        HVX_Vector vo = hvx_vec_splat_f32(offset);                                                                   \
+                                                                                                                     \
+        const uint32_t nvec = n / VLEN_FP16;                                                                         \
+        const uint32_t nloe = n % VLEN_FP16;                                                                         \
+                                                                                                                     \
+        uint32_t i = 0;                                                                                              \
+                                                                                                                     \
+        _Pragma("unroll(4)")                                                                                         \
+        for (; i < nvec; ++i) {                                                                                      \
+            HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);                                                          \
+            HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), vs), vo)); \
+            HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), vs), vo)); \
+            vdst[i] = hvx_vec_f32_to_f16(r0, r1);                                                                    \
+        }                                                                                                            \
+        if (nloe) {                                                                                                  \
+            HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);                                                          \
+            HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), vs), vo)); \
+            HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), vs), vo)); \
+            HVX_Vector v = hvx_vec_f32_to_f16(r0, r1);                                                               \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, v);                                                     \
+        }                                                                                                            \
+    } while(0)
+
+static inline void hvx_scale_offset_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
+    assert((size_t) dst % 128 == 0);
+    assert((size_t) src % 128 == 0);
+    hvx_scale_offset_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_scale_offset_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
+    assert((size_t) dst % 128 == 0);
+    hvx_scale_offset_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_scale_offset_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
+    assert((size_t) src % 128 == 0);
+    hvx_scale_offset_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_scale_offset_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
+    hvx_scale_offset_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_scale_offset_f16(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
+    if (((size_t) dst & 127) == 0) {
+        if (((size_t) src & 127) == 0) {
+            hvx_scale_offset_f16_aa(dst, src, n, scale, offset);
+        } else {
+            hvx_scale_offset_f16_au(dst, src, n, scale, offset);
+        }
+    } else {
+        if (((size_t) src & 127) == 0) {
+            hvx_scale_offset_f16_ua(dst, src, n, scale, offset);
+        } else {
+            hvx_scale_offset_f16_uu(dst, src, n, scale, offset);
+        }
+    }
+}
+
 #endif // HVX_SCALE_H

ggml/src/ggml-hexagon/htp/hvx-sqrt.h

@@ -123,4 +123,67 @@ static inline void hvx_sqrt_f32(uint8_t * restrict dst, const uint8_t * restrict
     }
 }
 
+// Compute sqrt(x) for f16 by promoting to f32, applying hvx_vec_rsqrt_f32, and narrowing back.
+#define hvx_sqrt_f16_loop_body(dst_type, src_type, vec_store)                          \
+    do {                                                                               \
+        dst_type * restrict vdst = (dst_type *) dst;                                   \
+        src_type * restrict vsrc = (src_type *) src;                                   \
+                                                                                       \
+        const uint32_t nvec = n / VLEN_FP16;                                           \
+        const uint32_t nloe = n % VLEN_FP16;                                           \
+                                                                                       \
+        uint32_t i = 0;                                                                \
+                                                                                       \
+        _Pragma("unroll(4)")                                                           \
+        for (; i < nvec; i++) {                                                        \
+            HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);                            \
+            HVX_Vector r0 = HVX_OP_MUL(hvx_vec_rsqrt_f32(Q6_V_lo_W(p)), Q6_V_lo_W(p)); \
+            HVX_Vector r1 = HVX_OP_MUL(hvx_vec_rsqrt_f32(Q6_V_hi_W(p)), Q6_V_hi_W(p)); \
+            vdst[i] = hvx_vec_f32_to_f16(r0, r1);                                      \
+        }                                                                              \
+        if (nloe) {                                                                    \
+            HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);                            \
+            HVX_Vector r0 = HVX_OP_MUL(hvx_vec_rsqrt_f32(Q6_V_lo_W(p)), Q6_V_lo_W(p)); \
+            HVX_Vector r1 = HVX_OP_MUL(hvx_vec_rsqrt_f32(Q6_V_hi_W(p)), Q6_V_hi_W(p)); \
+            HVX_Vector v = hvx_vec_f32_to_f16(r0, r1);                                 \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, v);                       \
+        }                                                                              \
+    } while(0)
+
+static inline void hvx_sqrt_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqrt_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqrt_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_sqrt_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqrt_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqrt_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqrt_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_sqrt_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqrt_f16(uint8_t * restrict dst, const uint8_t * restrict src, const int num_elems) {
+    if ((unsigned long) dst % 128 == 0) {
+        if ((unsigned long) src % 128 == 0) {
+            hvx_sqrt_f16_aa(dst, src, num_elems);
+        } else {
+            hvx_sqrt_f16_au(dst, src, num_elems);
+        }
+    } else {
+        if ((unsigned long) src % 128 == 0) {
+            hvx_sqrt_f16_ua(dst, src, num_elems);
+        } else {
+            hvx_sqrt_f16_uu(dst, src, num_elems);
+        }
+    }
+}
+
 #endif /* HVX_SQRT_H */

ggml/src/ggml-hexagon/htp/unary-ops.c

@@ -234,6 +234,146 @@ static void sqrt_f32(const float * restrict src,
     }
 }
 
+static void scale_f16(const _Float16 * restrict src,
+                      _Float16 * restrict dst,
+                      const uint32_t num_rows,
+                      const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float scale = 0.f;
+    float bias  = 0.f;
+    memcpy(&scale, &op_params[0], sizeof(float));
+    memcpy(&bias,  &op_params[1], sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_scale_offset_f16_aa((uint8_t *) dst_local, (const uint8_t *) src_local, ne0, scale, bias);
+    }
+}
+
+static void clamp_f16(const _Float16 * restrict src,
+                      _Float16 * restrict dst,
+                      const uint32_t num_rows,
+                      const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float min = 0.f;
+    float max = 0.f;
+    memcpy(&min, &op_params[0], sizeof(float));
+    memcpy(&max, &op_params[1], sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_clamp_scalar_f16(dst_local, src_local, (_Float16) min, (_Float16) max, ne0);
+    }
+}
+
+static void rms_norm_f16(const _Float16 * restrict src,
+                         _Float16 * restrict dst,
+                         const uint32_t num_rows,
+                         const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float epsilon = 0.f;
+    memcpy(&epsilon, op_params, sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_fast_rms_norm_f16((const uint8_t *) src_local, (uint8_t *) dst_local, ne0, epsilon);
+    }
+}
+
+static void norm_f16(const _Float16 * restrict src,
+                     _Float16 * restrict dst,
+                     const uint32_t num_rows,
+                     const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float epsilon = 0.f;
+    memcpy(&epsilon, op_params, sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_fast_norm_f16((const uint8_t *) src_local, (uint8_t *) dst_local, ne0, epsilon);
+    }
+}
+
+static void sqr_f16(const _Float16 * restrict src,
+                    _Float16 * restrict dst,
+                    const uint32_t num_rows,
+                    const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_sqr_f16_aa((uint8_t *) dst_local, (const uint8_t *) src_local, ne0);
+    }
+}
+
+static void sqrt_f16(const _Float16 * restrict src,
+                     _Float16 * restrict dst,
+                     const uint32_t num_rows,
+                     const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_sqrt_f16_aa((uint8_t *) dst_local, (const uint8_t *) src_local, ne0);
+    }
+}
+
+static void abs_f16(const _Float16 * restrict src,
+                    _Float16 * restrict dst,
+                    const uint32_t num_rows,
+                    const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_abs_f16_aa((uint8_t *) dst_local, (const uint8_t *) src_local, ne0);
+    }
+}
+
+static void log_f16(const _Float16 * restrict src,
+                    _Float16 * restrict dst,
+                    const uint32_t num_rows,
+                    const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_log_f16_aa((uint8_t *) dst_local, (const uint8_t *) src_local, ne0);
+    }
+}
+
+static void l2_norm_f16(const _Float16 * restrict src,
+                        _Float16 * restrict dst,
+                        const uint32_t num_rows,
+                        const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float epsilon = 0.f;
+    memcpy(&epsilon, op_params, sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_f = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_f       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_fast_l2_norm_f16((const uint8_t *)src_f, (uint8_t *)dst_f, ne0, epsilon);
+    }
+}
+
 static void neg_f32(const float * restrict src,
                     float * restrict dst,
                     const uint32_t num_rows,
@@ -471,8 +611,8 @@ static void log_f32(const float * restrict src,
     }
 }
 
-#define DEFINE_UNARY_TASK(NAME, IS_RMS_NORM_MUL, IS_TRI, CORE_EXPR)                                                 \
-static void unary_task_f32_##NAME(unsigned int nth, unsigned int ith, void * data) {                                \
+#define DEFINE_UNARY_TASK_IMPL(NAME, TYPE, SUFFIX, IS_RMS_NORM_MUL, IS_TRI, CORE_EXPR)                              \
+static void unary_task_##SUFFIX##_##NAME(unsigned int nth, unsigned int ith, void * data) {                         \
     const struct htp_unary_context * uctx = (const struct htp_unary_context *) data;                                \
     struct htp_ops_context * octx = uctx->octx;                                                                     \
     const struct htp_tensor * src = octx->src[0];                                                                   \
@@ -536,7 +676,7 @@ static void unary_task_f32_##NAME(unsigned int nth, unsigned int ith, void * dat
     const uint32_t dst_max_block  = block_dst_contig  ? uctx->block : MIN((uint32_t)uctx->block, ne1);              \
     const uint32_t BLOCK = MIN(src0_max_block, dst_max_block);                                                      \
     if (BLOCK == 0) {                                                                                               \
-        FARF(ERROR, "unary-f32 : current VTCM reservation %zu is too small, needed at least %zu\n",                 \
+        FARF(ERROR, "unary-" #SUFFIX " : current VTCM reservation %zu is too small, needed at least %zu\n",         \
              uctx->vtcm_src0_size_per_thread, src0_row_size_aligned);                                               \
         return;                                                                                                     \
     }                                                                                                               \
@@ -578,11 +718,11 @@ static void unary_task_f32_##NAME(unsigned int nth, unsigned int ith, void * dat
         const uint32_t block_size = unary_block_size(ir, src0_end_row, BLOCK, block_src0_contig, block_dst_contig,  \
                                                      ne01, div_ne01);                                               \
                                                                                                                     \
-        float * dst_vtcm  = (float *) dma_queue_pop(dma_queue).src;                                                 \
-        float * src0_vtcm = (float *) dma_queue_pop(dma_queue).dst;                                                 \
-        float * src1_vtcm = NULL;                                                                                   \
+        TYPE * dst_vtcm  = (TYPE *) dma_queue_pop(dma_queue).src;                                                   \
+        TYPE * src0_vtcm = (TYPE *) dma_queue_pop(dma_queue).dst;                                                   \
+        TYPE * src1_vtcm = NULL;                                                                                    \
         if ((IS_RMS_NORM_MUL) && !uctx->broadcast_weight) {                                                         \
-            src1_vtcm = (float *) dma_queue_pop(dma_queue).dst;                                                     \
+            src1_vtcm = (TYPE *) dma_queue_pop(dma_queue).dst;                                                      \
         }                                                                                                           \
                                                                                                                     \
         htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir);                                                      \
@@ -625,6 +765,10 @@ static void unary_task_f32_##NAME(unsigned int nth, unsigned int ith, void * dat
     dma_queue_flush(dma_queue);                                                                                     \
 }
 
+// F32 unary task: row-block DMA/VTCM plumbing, float-typed VTCM buffers.
+#define DEFINE_UNARY_TASK(NAME, IS_RMS_NORM_MUL, IS_TRI, CORE_EXPR) \
+    DEFINE_UNARY_TASK_IMPL(NAME, float, f32, IS_RMS_NORM_MUL, IS_TRI, CORE_EXPR)
+
 DEFINE_UNARY_TASK(norm,           false, false, norm_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(rms_norm,       false, false, rms_norm_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(rms_norm_mul,   true,  false, rms_norm_mul_f32(src0_vtcm, uctx->broadcast_weight ? (const float *) src1_vtcm_data : src1_vtcm, dst_vtcm, block_size, uctx))
@@ -644,6 +788,18 @@ DEFINE_UNARY_TASK(unary_log,      false, false, log_f32(src0_vtcm, dst_vtcm, blo
 DEFINE_UNARY_TASK(l2_norm,        false, false, l2_norm_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(tri,            false, true,  tri_f32(src0_vtcm, dst_vtcm, block_size, ir, uctx))
 
+// F16 unary tasks: same DMA/VTCM plumbing as DEFINE_UNARY_TASK, but VTCM buffers are
+// _Float16-typed. None of the current F16 ops need RMS_NORM_MUL or TRI support.
+DEFINE_UNARY_TASK_IMPL(norm,      _Float16, f16, false, false, norm_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(rms_norm,  _Float16, f16, false, false, rms_norm_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(scale,     _Float16, f16, false, false, scale_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(clamp,     _Float16, f16, false, false, clamp_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(sqr,       _Float16, f16, false, false, sqr_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(sqrt,      _Float16, f16, false, false, sqrt_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(l2_norm,   _Float16, f16, false, false, l2_norm_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(unary_abs, _Float16, f16, false, false, abs_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(unary_log, _Float16, f16, false, false, log_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+
 // Apply a pointwise unary op to one column tile that is already in VTCM.
 #define DEFINE_UNARY_TILED_TASK(NAME, IS_TRI, CORE_TILE_EXPR)                                                       \
 static void unary_task_f32_tiled_##NAME(unsigned int nth, unsigned int ith, void * data) {                          \
@@ -892,50 +1048,76 @@ DEFINE_UNARY_TILED_TASK(unary_abs,      false, hvx_abs_f32_aa(dst_vtcm, src_vtcm
 DEFINE_UNARY_TILED_TASK(unary_log,      false, hvx_log_f32_aa(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(tri,            true,  tri_apply_tile_f32(src_vtcm, dst_vtcm, tw, col, i01, ne0, tri_ttype))
 
-static int execute_op_unary_f32(struct htp_ops_context * octx) {
+static int execute_op_unary(struct htp_ops_context * octx) {
     int err = HTP_STATUS_OK;
 
     const struct htp_tensor * src0 = octx->src[0];
     const struct htp_tensor * dst  = octx->dst;
 
+    const bool is_f16 = (src0->type == HTP_TYPE_F16);
+
     const char * op_type = NULL;
 
     switch (octx->op) {
-        case HTP_OP_NORM:            op_type = "norm-f32";         break;
-        case HTP_OP_RMS_NORM:        op_type = "rmsnorm-f32";      break;
-        case HTP_OP_RMS_NORM_MUL:    op_type = "rmsnorm-mul-f32";  break;
-        case HTP_OP_SCALE:           op_type = "scale-f32";        break;
-        case HTP_OP_CLAMP:           op_type = "clamp-f32";        break;
-        case HTP_OP_SQR:             op_type = "sqr-f32";          break;
-        case HTP_OP_SQRT:            op_type = "sqrt-f32";         break;
-        case HTP_OP_UNARY_NEG:       op_type = "neg-f32";          break;
-        case HTP_OP_UNARY_EXP:       op_type = "exp-f32";          break;
-        case HTP_OP_UNARY_SIGMOID:   op_type = "sigmoid-f32";      break;
-        case HTP_OP_UNARY_SILU:      op_type = "silu-f32";         break;
-        case HTP_OP_UNARY_GELU:      op_type = "gelu-f32";         break;
-        case HTP_OP_UNARY_SOFTPLUS:  op_type = "softplus-f32";     break;
-        case HTP_OP_UNARY_TANH:      op_type = "tanh-f32";         break;
-        case HTP_OP_UNARY_ABS:       op_type = "abs-f32";          break;
-        case HTP_OP_UNARY_LOG:       op_type = "log-f32";          break;
-        case HTP_OP_L2_NORM:         op_type = "l2norm-f32";       break;
-        case HTP_OP_TRI:             op_type = "tri-f32";          break;
+        case HTP_OP_NORM:            op_type = is_f16 ? "norm-f16"     : "norm-f32";         break;
+        case HTP_OP_RMS_NORM:        op_type = is_f16 ? "rmsnorm-f16"  : "rmsnorm-f32";      break;
+        case HTP_OP_RMS_NORM_MUL:    op_type = "rmsnorm-mul-f32";                            break;
+        case HTP_OP_SCALE:           op_type = is_f16 ? "scale-f16"    : "scale-f32";        break;
+        case HTP_OP_CLAMP:           op_type = is_f16 ? "clamp-f16"    : "clamp-f32";        break;
+        case HTP_OP_SQR:             op_type = is_f16 ? "sqr-f16"      : "sqr-f32";          break;
+        case HTP_OP_SQRT:            op_type = is_f16 ? "sqrt-f16"     : "sqrt-f32";         break;
+        case HTP_OP_UNARY_NEG:       op_type = "neg-f32";                                    break;
+        case HTP_OP_UNARY_EXP:       op_type = "exp-f32";                                    break;
+        case HTP_OP_UNARY_SIGMOID:   op_type = "sigmoid-f32";                                break;
+        case HTP_OP_UNARY_SILU:      op_type = "silu-f32";                                   break;
+        case HTP_OP_UNARY_GELU:      op_type = "gelu-f32";                                   break;
+        case HTP_OP_UNARY_SOFTPLUS:  op_type = "softplus-f32";                               break;
+        case HTP_OP_UNARY_TANH:      op_type = "tanh-f32";                                   break;
+        case HTP_OP_UNARY_ABS:       op_type = is_f16 ? "abs-f16"      : "abs-f32";          break;
+        case HTP_OP_UNARY_LOG:       op_type = is_f16 ? "log-f16"      : "log-f32";          break;
+        case HTP_OP_L2_NORM:         op_type = is_f16 ? "l2norm-f16"   : "l2norm-f32";       break;
+        case HTP_OP_TRI:             op_type = "tri-f32";                                    break;
 
         default:
             FARF(ERROR, "Unsupported unary Op %u\n", octx->op);
             return HTP_STATUS_NO_SUPPORT;
     }
 
+    // F16 only has row-block kernels for this subset of ops (see the dispatch switch
+    // below) - reject everything else up front, before touching kparams/VTCM.
+    if (is_f16) {
+        switch (octx->op) {
+            case HTP_OP_NORM:
+            case HTP_OP_RMS_NORM:
+            case HTP_OP_SCALE:
+            case HTP_OP_CLAMP:
+            case HTP_OP_SQR:
+            case HTP_OP_SQRT:
+            case HTP_OP_L2_NORM:
+            case HTP_OP_UNARY_ABS:
+            case HTP_OP_UNARY_LOG:
+                break;
+            default:
+                FARF(ERROR, "unary-%s: not supported for F16\n", op_type);
+                return HTP_STATUS_NO_SUPPORT;
+        }
+    }
+
     const struct htp_unary_kernel_params * kparams = (const struct htp_unary_kernel_params *) octx->kernel_params;
 
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
     const uint32_t n_threads  = kparams->n_threads;
 
-    const size_t src0_data_row_size = src0->ne[0] * sizeof(float);
-    const size_t dst_data_row_size  = dst->ne[0]  * sizeof(float);
+    const size_t elem_size = is_f16 ? sizeof(_Float16) : sizeof(float);
+
+    const size_t src0_data_row_size = src0->ne[0] * elem_size;
+    const size_t dst_data_row_size  = dst->ne[0]  * elem_size;
 
     const size_t src0_row_size_aligned = kparams->src0_row_size_aligned;
     const size_t dst_row_size_aligned  = kparams->dst_row_size_aligned;
 
+    // Always 0 for F16 - htp_unary_vtcm_layout_build() keeps F16 on the row-block path,
+    // since only F32 has unary_task_f32_tiled_* kernels.
     const uint32_t col_tile = kparams->col_tile;
 
     size_t src1_data_row_size = 0;
@@ -943,6 +1125,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
     bool broadcast_weight = kparams->broadcast_weight;
     const struct htp_tensor * src1 = NULL;
 
+    // RMS_NORM_MUL fusion is F32-only (its weight tensor is always F32; see
+    // try_fuse_node()'s type guard), so this never triggers when is_f16 is true.
     if (octx->op == HTP_OP_RMS_NORM_MUL) {
         src1 = octx->src[1];
         src1_data_row_size = src1->ne[0] * sizeof(float);
@@ -987,7 +1171,7 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
 
             .block                 = kparams->block,
             .nc                    = src0->ne[0],
-            .col_tile              = (uint32_t) kparams->col_tile,
+            .col_tile              = col_tile,
             .broadcast_weight      = broadcast_weight,
 
             .vtcm_src0             = VTCM_LAYOUT_PTR(uint8_t, base, 0),
@@ -1020,6 +1204,19 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
                 case HTP_OP_TRI:             task_func = unary_task_f32_tiled_tri;            break;
                 default:                     break;
             }
+        } else if (is_f16) {
+            switch (octx->op) {
+                case HTP_OP_NORM:            task_func = unary_task_f16_norm;                 break;
+                case HTP_OP_RMS_NORM:        task_func = unary_task_f16_rms_norm;             break;
+                case HTP_OP_SCALE:           task_func = unary_task_f16_scale;                break;
+                case HTP_OP_CLAMP:           task_func = unary_task_f16_clamp;                break;
+                case HTP_OP_SQR:             task_func = unary_task_f16_sqr;                  break;
+                case HTP_OP_SQRT:            task_func = unary_task_f16_sqrt;                 break;
+                case HTP_OP_L2_NORM:         task_func = unary_task_f16_l2_norm;              break;
+                case HTP_OP_UNARY_ABS:       task_func = unary_task_f16_unary_abs;            break;
+                case HTP_OP_UNARY_LOG:       task_func = unary_task_f16_unary_log;            break;
+                default:                     break;
+            }
         } else {
             switch (octx->op) {
                 case HTP_OP_NORM:            task_func = unary_task_f32_norm;                 break;
@@ -1047,7 +1244,7 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
         if (task_func) {
             worker_pool_run_func(octx->ctx->worker_pool, task_func, &uctx, n_threads);
         } else {
-            FARF(ERROR, "execute_op_unary_f32: task function is NULL for op %d\n", octx->op);
+            FARF(ERROR, "execute_op_unary: task function is NULL for op %d\n", octx->op);
             err = HTP_STATUS_NO_SUPPORT;
         }
     }
@@ -1058,7 +1255,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
 int op_unary(struct htp_ops_context * octx) {
     switch (octx->src[0]->type) {
         case HTP_TYPE_F32:
-            return execute_op_unary_f32(octx);
+        case HTP_TYPE_F16:
+            return execute_op_unary(octx);
 
         default:
             return HTP_STATUS_NO_SUPPORT;

ggml/src/ggml-hexagon/htp/unary-ops.h

@@ -85,17 +85,19 @@ static inline void htp_unary_vtcm_layout_build(
     bool broadcast_weight,
     uint32_t n_threads,
     size_t vtcm_size,
+    size_t elem_size,
     uint32_t * out_col_tile,
     uint32_t * out_vtcm_row_per_thread
 ) {
-    const size_t src0_data_row_size = ne00 * sizeof(float);
-    const size_t dst_data_row_size  = ne10 * sizeof(float);
+    const size_t src0_data_row_size = ne00 * elem_size;
+    const size_t dst_data_row_size  = ne10 * elem_size;
 
     const size_t src0_row_size_aligned = hex_round_up(src0_data_row_size, 128);
     const size_t dst_row_size_aligned  = hex_round_up(dst_data_row_size,  128);
 
     size_t src1_row_size_aligned = 0;
     if (op == HTP_OP_RMS_NORM_MUL) {
+        // RMS_NORM_MUL fusion is F32-only; its weight tensor is always F32.
         const size_t src1_data_row_size = ne11 * sizeof(float);
         src1_row_size_aligned = hex_round_up(src1_data_row_size, 128);
     }
@@ -125,12 +127,19 @@ static inline void htp_unary_vtcm_layout_build(
 
     const bool is_reduction = (op == HTP_OP_NORM || op == HTP_OP_RMS_NORM ||
                                op == HTP_OP_RMS_NORM_MUL || op == HTP_OP_L2_NORM);
+    // The tiled fallback path below only has F32 task functions (unary_task_f32_tiled_*);
+    // F16 has no tiled kernels, so it must stay on the row-block path like reduction ops.
+    // NOTE: if F16 ends up with vtcm_row_per_thread == 0 here (row too large for the VTCM
+    // budget), execute_op_unary() will see BLOCK == 0 and skip computation for that op
+    // (logged via FARF(ERROR, ...)) since there is no F16 tiled fallback. This is a known
+    // limitation; supporting it would require adding F16 tiled kernels.
+    const bool is_f16 = (elem_size == sizeof(_Float16));
     uint32_t col_tile = 0;
 
-    if (vtcm_row_per_thread == 0 && !is_reduction) {
+    if (vtcm_row_per_thread == 0 && !is_reduction && !is_f16) {
         const size_t per_thread_budget = vtcm_size / n_threads;
         const size_t col_tile_bytes = hex_align_down(per_thread_budget / 4, 128);
-        col_tile = (uint32_t) (col_tile_bytes / sizeof(float));
+        col_tile = (uint32_t) (col_tile_bytes / elem_size);
 
         L->src0_bytes = col_tile_bytes * 2;
         L->dst_bytes  = col_tile_bytes * 2;

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

@@ -1471,8 +1471,10 @@ void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_eve
 
 void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total) {
     if (@available(macOS 10.12, iOS 16.0, *)) {
-        *total = dev->mtl_device.recommendedMaxWorkingSetSize;
-        *free  = *total - dev->mtl_device.currentAllocatedSize;
+        *total     = dev->mtl_device.recommendedMaxWorkingSetSize;
+        size_t cur = dev->mtl_device.currentAllocatedSize;
+        // it's possible to allocate more than `recommendedMaxWorkingSetSize`
+        *free      = *total > cur ? *total - cur : 0;
     } else {
         *free = 0;
         *total = 0;

ggml/src/ggml-metal/ggml-metal-tuning.cpp

@@ -1468,6 +1468,107 @@ constexpr fa_vec_entry_t fa_vec_tuned_table[] = {
     { { GGML_METAL_DEVICE_M2_ULTRA, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } },
     { { GGML_METAL_DEVICE_M2_ULTRA, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },
 
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 1, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 2, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, 2, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 128, 128, 1, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 128, 128, 2, 3 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 128, 128, 3, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 128, 128, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 192, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 192, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 192, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 192, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 192, 2, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 2, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 3, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 3, 2 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 3, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 1, 1 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 1, 2 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 1, 4 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 320, 256, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 320, 256, 3, 0 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 320, 256, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 320, 256, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 512, 512, 2, 0 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 512, 512, 3, 0 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 512, 512, 3, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 512, 512, 3, 3 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 2, 0 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 2, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 2, 2 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 3, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 96, 96, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, 3, 4 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, 2, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 128, 1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 128, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 256, 256, 3, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 320, 256, 1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 320, 256, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 320, 256, 1, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },
+
     { { GGML_METAL_DEVICE_M3_PRO, GGML_TYPE_F16, 32, 32, 1, 1 }, { 2, 4 } },
     { { GGML_METAL_DEVICE_M3_PRO, GGML_TYPE_F16, 32, 32, 1, 3 }, { 2, 4 } },
     { { GGML_METAL_DEVICE_M3_PRO, GGML_TYPE_F16, 32, 32, 2, 1 }, { 2, 4 } },

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

@@ -742,6 +742,7 @@ static void dev2dev_memcpy(int device_dst, sycl::queue &q_dst, int device_src, s
         if (q_dst.get_device().ext_oneapi_can_access_peer(q_src.get_device(),
                                                           sycl::ext::oneapi::peer_access::access_supported)) {
             GGML_SYCL_DEBUG("[SYCL] dev2dev memcpy by SYCL\n");
+            q_dst.get_device().ext_oneapi_enable_peer_access(q_src.get_device());
             SYCL_CHECK(CHECK_TRY_ERROR(q_dst.memcpy(ptr_dst, ptr_src, size).wait()));
             return;
         }

ggml/src/ggml-sycl/mmvq.cpp

@@ -6,6 +6,24 @@
 #include "quants.hpp"
 #include "vecdotq.hpp"
 
+// Minimum weight-row count at which the Q4_K multi-column MMVQ kernel handles two output rows per
+// subgroup (rows_per_sg == 2) instead of one, when ncols_dst == 2.
+//
+// Pairing rows lets a subgroup load each activation block once and apply it to two rows, at the cost
+// of halving the number of subgroups in the launch. With only two destination columns there is too
+// little work per row to hide that loss of parallelism, so pairing only pays off once there are
+// enough rows to keep the device occupied. This is a measured performance crossover, not a
+// correctness or hardware limit - both variants compute the same result for any nrows.
+//
+// Derived on Intel Arc Pro B70 with `test-backend-ops perf -o MUL_MAT` (Q4_K, ncols_dst == 2),
+// sweeping nrows over 5120..6912 at ncols 17408 and 19968: one row per subgroup was up to 9% faster
+// below the crossover, two rows per subgroup 8-15% faster above it, and the crossover fell inside
+// (6144, 6272] for both ncols with no measurable ncols dependence. A later 32-row granularity sweep
+// narrowed it to (6144, 6176], so 6272 is a conservative gate rather than the exact crossover.
+// ncols_dst >= 3 amortizes the activation loads over more columns and is faster with two rows at
+// every row count, so it does not consult this threshold.
+static constexpr int Q4_K_MMVQ_ROW_PAIR_MIN_NROWS = 6272;
+
 template <typename reorder_vec_dot_q_sycl>
 static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
                                   const int ncols, const int nrows, const sycl::nd_item<3> & nd_item) {
@@ -59,7 +77,7 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
 
 // With has_fusion, `vgate` is a second weight matrix sharing vx's shape, stride and reorder
 // layout: one pass computes both row dot products and the epilogue writes glu(gate, up).
-template <typename reorder_vec_dot_q_sycl, int ncols_dst, bool has_fusion = false>
+template <typename reorder_vec_dot_q_sycl, int ncols_dst, bool has_fusion = false, int rows_per_sg = 1>
 static void mul_mat_vec_q_reorder_ncols(const void * __restrict__ vx, const void * __restrict__ vgate,
                                         const void * __restrict__ vy, float * __restrict__ dst, const int ncols,
                                         const int nrows, const int stride_col_y_bytes, const int stride_col_dst,
@@ -71,14 +89,17 @@ static void mul_mat_vec_q_reorder_ncols(const void * __restrict__ vx, const void
     const int  sg_range     = sg.get_group_linear_range();
     const int  workgroup_id = nd_item.get_group_linear_id();
     const int  sg_id        = sg.get_group_linear_id();
-    const int  row          = workgroup_id * sg_range + sg_id;
+    const int  row0         = (workgroup_id * sg_range + sg_id) * rows_per_sg;
 
     // row is sub-group uniform, so this retires whole sub-groups and the collectives below
     // stay convergent
-    if (row >= nrows) {
+    if (row0 >= nrows) {
         return;
     }
 
+    static_assert(rows_per_sg == 1 ||
+                  reorder_vec_dot_shared_activations<reorder_vec_dot_q_sycl::gtype>::value);
+
     const int     blocks_per_row              = ncols / block_traits::qk;
     constexpr int blocks_per_subgroup         = ceil_div(block_traits::vdr_mmvq * WARP_SIZE, block_traits::qi);
     constexpr int block_elements_per_subgroup = block_traits::qi / block_traits::vdr_mmvq;
@@ -87,34 +108,96 @@ static void mul_mat_vec_q_reorder_ncols(const void * __restrict__ vx, const void
     static_assert(blocks_per_subgroup > 0);
     static_assert(block_elements_per_subgroup > 0);
 
-    float partial_sum[ncols_dst] = { 0.0f };
+    float partial_sum[ncols_dst][rows_per_sg] = {};
     // sized 1 rather than 0 when unused: zero-length arrays are not standard C++, and the
     // array is dead and eliminated in that case
-    [[maybe_unused]] float partial_gate[has_fusion ? ncols_dst : 1] = { 0.0f };
+    [[maybe_unused]] float partial_gate[has_fusion ? ncols_dst : 1][has_fusion ? rows_per_sg : 1] = {};
     for (int i = sg.get_local_linear_id() / block_elements_per_subgroup; i < blocks_per_row; i += blocks_per_subgroup) {
-        const int ibx = row * blocks_per_row + i;
-
-        // the offsets depend only on the block index and the matrix shape, never on the base
-        // pointer, which is what lets vgate reuse them
-        const auto bx_offset = block_type::get_block_offset(ibx, nblocks);
-        const auto d_offset  = block_type::get_d_offset(nrows, ncols, ibx);
         const int  iby       = i * block_type::block_to_q8_1_ratio();
 
 #pragma unroll
         for (int elem = 0; elem < block_elements_per_subgroup; elem += WARP_SIZE) {
             const int iqs = elem + block_traits::vdr_mmvq * (sg.get_local_linear_id() % block_elements_per_subgroup);
 
+            if constexpr (rows_per_sg > 1) {
+                typename reorder_vec_dot_q_sycl::weights wx[rows_per_sg];
+                [[maybe_unused]] typename reorder_vec_dot_q_sycl::weights wg[rows_per_sg];
 #pragma unroll
-            for (int j = 0; j < ncols_dst; ++j) {
-                const char        * vy_j           = (const char *) vy + j * stride_col_y_bytes;
-                const int8_t      * q8_1_quant_ptr = (const int8_t *) vy_j + iby * QK8_1;
-                const sycl::half2 * q8_1_ds_ptr    = (const sycl::half2 *) (vy_j + ncols + iby * sizeof(sycl::half2));
-
-                partial_sum[j] += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, q8_1_quant_ptr, q8_1_ds_ptr, iqs);
-
+                for (int r = 0; r < rows_per_sg; ++r) {
+                    const int row = sycl::min(row0 + r, nrows - 1);
+                    const int ibx = row * blocks_per_row + i;
+                    const auto bx_offset = block_type::get_block_offset(ibx, nblocks);
+                    const auto d_offset  = block_type::get_d_offset(nrows, ncols, ibx);
+                    wx[r] = reorder_vec_dot_q_sycl::load(vx, bx_offset, d_offset, iqs);
+                    if constexpr (has_fusion) {
+                        wg[r] = reorder_vec_dot_q_sycl::load(vgate, bx_offset, d_offset, iqs);
+                    }
+                }
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    const char        * vy_j           = (const char *) vy + j * stride_col_y_bytes;
+                    const int8_t      * q8_1_quant_ptr = (const int8_t *) vy_j + iby * QK8_1;
+                    const sycl::half2 * q8_1_ds_ptr =
+                        (const sycl::half2 *) (vy_j + ncols + iby * sizeof(sycl::half2));
+                    const auto a = reorder_vec_dot_q_sycl::load_activations(q8_1_quant_ptr, q8_1_ds_ptr, iqs);
+#pragma unroll
+                    for (int r = 0; r < rows_per_sg; ++r) {
+                        partial_sum[j][r] += reorder_vec_dot_q_sycl::apply(wx[r], a);
+                        if constexpr (has_fusion) {
+                            partial_gate[j][r] += reorder_vec_dot_q_sycl::apply(wg[r], a);
+                        }
+                    }
+                }
+            } else if constexpr (reorder_vec_dot_shared_weights<reorder_vec_dot_q_sycl::gtype>::value) {
+                const int ibx = row0 * blocks_per_row + i;
+                const auto bx_offset = block_type::get_block_offset(ibx, nblocks);
+                const auto d_offset  = block_type::get_d_offset(nrows, ncols, ibx);
+                const auto wx = reorder_vec_dot_q_sycl::load(vx, bx_offset, d_offset, iqs);
                 if constexpr (has_fusion) {
-                    partial_gate[j] +=
-                        reorder_vec_dot_q_sycl()(vgate, bx_offset, d_offset, q8_1_quant_ptr, q8_1_ds_ptr, iqs);
+                    const auto wg = reorder_vec_dot_q_sycl::load(vgate, bx_offset, d_offset, iqs);
+
+#pragma unroll
+                    for (int j = 0; j < ncols_dst; ++j) {
+                        const char        * vy_j           = (const char *) vy + j * stride_col_y_bytes;
+                        const int8_t      * q8_1_quant_ptr = (const int8_t *) vy_j + iby * QK8_1;
+                        const sycl::half2 * q8_1_ds_ptr =
+                            (const sycl::half2 *) (vy_j + ncols + iby * sizeof(sycl::half2));
+
+                        // up and gate share the activation, so load it once and apply it twice
+                        const auto a = reorder_vec_dot_q_sycl::load_activations(q8_1_quant_ptr, q8_1_ds_ptr, iqs);
+
+                        partial_sum[j][0] += reorder_vec_dot_q_sycl::apply(wx, a);
+                        partial_gate[j][0] += reorder_vec_dot_q_sycl::apply(wg, a);
+                    }
+                } else {
+#pragma unroll
+                    for (int j = 0; j < ncols_dst; ++j) {
+                        const char        * vy_j           = (const char *) vy + j * stride_col_y_bytes;
+                        const int8_t      * q8_1_quant_ptr = (const int8_t *) vy_j + iby * QK8_1;
+                        const sycl::half2 * q8_1_ds_ptr =
+                            (const sycl::half2 *) (vy_j + ncols + iby * sizeof(sycl::half2));
+
+                        partial_sum[j][0] += reorder_vec_dot_q_sycl::dot(wx, q8_1_quant_ptr, q8_1_ds_ptr, iqs);
+                    }
+                }
+            } else {
+                const int ibx = row0 * blocks_per_row + i;
+                const auto bx_offset = block_type::get_block_offset(ibx, nblocks);
+                const auto d_offset  = block_type::get_d_offset(nrows, ncols, ibx);
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    const char        * vy_j           = (const char *) vy + j * stride_col_y_bytes;
+                    const int8_t      * q8_1_quant_ptr = (const int8_t *) vy_j + iby * QK8_1;
+                    const sycl::half2 * q8_1_ds_ptr =
+                        (const sycl::half2 *) (vy_j + ncols + iby * sizeof(sycl::half2));
+
+                    partial_sum[j][0] +=
+                        reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, q8_1_quant_ptr, q8_1_ds_ptr, iqs);
+
+                    if constexpr (has_fusion) {
+                        partial_gate[j][0] +=
+                            reorder_vec_dot_q_sycl()(vgate, bx_offset, d_offset, q8_1_quant_ptr, q8_1_ds_ptr, iqs);
+                    }
                 }
             }
         }
@@ -122,17 +205,20 @@ static void mul_mat_vec_q_reorder_ncols(const void * __restrict__ vx, const void
 
 #pragma unroll
     for (int j = 0; j < ncols_dst; ++j) {
-        float sum = sycl::reduce_over_group(nd_item.get_sub_group(), partial_sum[j], std::plus<>());
+#pragma unroll
+        for (int r = 0; r < rows_per_sg; ++r) {
+            float sum = sycl::reduce_over_group(nd_item.get_sub_group(), partial_sum[j][r], std::plus<>());
 
-        if constexpr (has_fusion) {
-            const float gate = sycl::reduce_over_group(nd_item.get_sub_group(), partial_gate[j], std::plus<>());
+            if constexpr (has_fusion) {
+                const float gate = sycl::reduce_over_group(nd_item.get_sub_group(), partial_gate[j][r], std::plus<>());
 
-            // uniform across the launch; the launcher only instantiates SWIGLU and GEGLU
-            sum *= glu_op == GGML_GLU_OP_SWIGLU ? op_silu(gate) : op_gelu(gate);
-        }
+                // uniform across the launch; the launcher only instantiates SWIGLU and GEGLU
+                sum *= glu_op == GGML_GLU_OP_SWIGLU ? op_silu(gate) : op_gelu(gate);
+            }
 
-        if (sg.leader()) {
-            dst[j * stride_col_dst + row] = sum;
+            if (sg.leader() && row0 + r < nrows) {
+                dst[j * stride_col_dst + row0 + r] = sum;
+            }
         }
     }
 }
@@ -1671,8 +1757,8 @@ static void reorder_mul_mat_vec_q4_k_q8_1_sycl(const void * vx, const void * vy,
     });
 }
 
-template <int ncols_dst>
-static void reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols(
+template <int ncols_dst, int rows_per_sg>
+static void reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols_impl(
         const void * vx, const void * vy, float * dst,
         const int ncols, const int nrows,
         const int stride_col_y_bytes, const int stride_col_dst,
@@ -1680,20 +1766,31 @@ static void reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols(
     GGML_ASSERT(ncols % QK_K == 0);
 
     constexpr size_t num_subgroups = WARP_SIZE;
-    const int block_num_y = ceil_div(nrows, GGML_SYCL_MMV_Y * (int) num_subgroups);
+    const int block_num_y = ceil_div(nrows, GGML_SYCL_MMV_Y * (int) num_subgroups * rows_per_sg);
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
 
     stream->submit([&](sycl::handler & cgh) {
         cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                             mul_mat_vec_q_reorder_ncols<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>, ncols_dst>(
+                             mul_mat_vec_q_reorder_ncols<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>, ncols_dst,
+                                                        /*has_fusion=*/ false, rows_per_sg>(
                                  vx, /*vgate=*/ nullptr, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst,
                                  /*glu_op=*/ GGML_GLU_OP_SWIGLU, nd_item);
                          });
     });
 }
 
+template <int ncols_dst>
+static void reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols(
+        const void * vx, const void * vy, float * dst,
+        const int ncols, const int nrows,
+        const int stride_col_y_bytes, const int stride_col_dst,
+        dpct::queue_ptr stream) {
+    constexpr int rows_per_sg = ncols_dst >= 3 && ncols_dst <= 4 ? 2 : 1;
+    reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols_impl<ncols_dst, rows_per_sg>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream);
+}
+
 static void reorder_mul_mat_vec_q4_k_q8_1_sycl_switch_ncols(
         const void * vx, const void * vy, float * dst,
         const int ncols, const int nrows, const int ncols_dst,
@@ -1701,7 +1798,13 @@ static void reorder_mul_mat_vec_q4_k_q8_1_sycl_switch_ncols(
         dpct::queue_ptr stream) {
     switch (ncols_dst) {
         case 1: reorder_mul_mat_vec_q4_k_q8_1_sycl(vx, vy, dst, ncols, nrows, stream); break;
-        case 2: reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols<2>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
+        case 2:
+            if (nrows >= Q4_K_MMVQ_ROW_PAIR_MIN_NROWS) {
+                reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols_impl<2, 2>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream);
+            } else {
+                reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols_impl<2, 1>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream);
+            }
+            break;
         case 3: reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols<3>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
         case 4: reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols<4>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
         case 5: reorder_mul_mat_vec_q4_k_q8_1_sycl_ncols<5>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
@@ -2839,8 +2942,8 @@ bool ggml_sycl_mul_mat_vec_q_id_reorder(
     }
 }
 
-template <typename reorder_vec_dot_q_sycl, int ncols_dst>
-static void launch_mul_mat_vec_q_reorder_glu(const void * vx, const void * vgate, const void * vy, float * dst,
+template <typename reorder_vec_dot_q_sycl, int ncols_dst, int rows_per_sg>
+static void launch_mul_mat_vec_q_reorder_glu_impl(const void * vx, const void * vgate, const void * vy, float * dst,
                                              const int ncols, const int nrows, const int stride_col_y_bytes,
                                              const int stride_col_dst, const ggml_glu_op glu_op,
                                              dpct::queue_ptr stream) {
@@ -2848,20 +2951,33 @@ static void launch_mul_mat_vec_q_reorder_glu(const void * vx, const void * vgate
 
     constexpr size_t num_subgroups = WARP_SIZE;
 
-    const int            block_num_y = ceil_div(nrows, GGML_SYCL_MMV_Y * (int) num_subgroups);
+    const int            block_num_y = ceil_div(nrows, GGML_SYCL_MMV_Y * (int) num_subgroups * rows_per_sg);
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
 
     stream->submit([&](sycl::handler & cgh) {
         cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                             mul_mat_vec_q_reorder_ncols<reorder_vec_dot_q_sycl, ncols_dst, /*has_fusion=*/ true>(
+                             mul_mat_vec_q_reorder_ncols<reorder_vec_dot_q_sycl, ncols_dst, /*has_fusion=*/ true,
+                                                        rows_per_sg>(
                                  vx, vgate, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, glu_op,
                                  nd_item);
                          });
     });
 }
 
+template <typename reorder_vec_dot_q_sycl, int ncols_dst>
+static void launch_mul_mat_vec_q_reorder_glu(const void * vx, const void * vgate, const void * vy, float * dst,
+                                             const int ncols, const int nrows, const int stride_col_y_bytes,
+                                             const int stride_col_dst, const ggml_glu_op glu_op,
+                                             dpct::queue_ptr stream) {
+    constexpr int rows_per_sg =
+        reorder_vec_dot_shared_activations<reorder_vec_dot_q_sycl::gtype>::value && ncols_dst >= 3 && ncols_dst <= 4
+            ? 2
+            : 1;
+    launch_mul_mat_vec_q_reorder_glu_impl<reorder_vec_dot_q_sycl, ncols_dst, rows_per_sg>(vx, vgate, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, glu_op, stream);
+}
+
 bool ggml_sycl_mul_mat_vec_q_glu_reorder(enum ggml_type src0_type, enum ggml_glu_op glu_op, const void * vx,
                                          const void * vgate, const void * vy, float * dst, int ncols, int nrows,
                                          int ncols_dst, int stride_col_y_bytes, int stride_col_dst,
@@ -2881,8 +2997,11 @@ bool ggml_sycl_mul_mat_vec_q_glu_reorder(enum ggml_type src0_type, enum ggml_glu
                                                          stride_col_dst, glu_op, stream);
             return true;
         case 2:
-            launch_mul_mat_vec_q_reorder_glu<vec_dot, 2>(vx, vgate, vy, dst, ncols, nrows, stride_col_y_bytes,
-                                                         stride_col_dst, glu_op, stream);
+            if (nrows >= Q4_K_MMVQ_ROW_PAIR_MIN_NROWS) {
+                launch_mul_mat_vec_q_reorder_glu_impl<vec_dot, 2, 2>(vx, vgate, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, glu_op, stream);
+            } else {
+                launch_mul_mat_vec_q_reorder_glu_impl<vec_dot, 2, 1>(vx, vgate, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, glu_op, stream);
+            }
             return true;
         case 3:
             launch_mul_mat_vec_q_reorder_glu<vec_dot, 3>(vx, vgate, vy, dst, ncols, nrows, stride_col_y_bytes,

ggml/src/ggml-sycl/vecdotq.hpp

@@ -351,6 +351,25 @@ template <ggml_type T> struct reorder_vec_dot_q_sycl {
     static_assert(T != T, "ggml_type for reorder vecdot not implemented");
 };
 
+// For some types the weight side of the dot product does not depend on the destination column, so a
+// multi-column mul_mat_vec can unpack it once per block instead of once per column. Such a type adds
+// load() and dot() next to operator() and opts in here. See reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>.
+template <ggml_type T> struct reorder_vec_dot_shared_weights {
+    static constexpr bool value = false;
+};
+
+template <> struct reorder_vec_dot_shared_weights<GGML_TYPE_Q4_K> {
+    static constexpr bool value = true;
+};
+
+template <ggml_type T> struct reorder_vec_dot_shared_activations {
+    static constexpr bool value = false;
+};
+
+template <> struct reorder_vec_dot_shared_activations<GGML_TYPE_Q4_K> {
+    static constexpr bool value = true;
+};
+
 template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
     static constexpr ggml_type gtype = GGML_TYPE_Q4_0;
 
@@ -540,50 +559,84 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
     using q4_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q4_K>;
     using q4_k_traits = typename q4_k_block::traits;
 
+    struct weights {
+        int        v[2];
+        uint16_t   aux[2];
+        ggml_half2 dm;
+        int        bq8_offset;
+    };
+
+    struct activations {
+        int   u[2 * QR4_K];
+        float d8[QR4_K];
+    };
+
+    __dpct_inline__ static weights load(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                        const std::pair<int, int> d_offset, const int & iqs) {
+        const uint8_t *    base = static_cast<const uint8_t *>(vbq);
+        const uint8_t *    qs   = base + ibx_offset.first;
+        const uint8_t *    scs  = base + d_offset.first;
+        const ggml_half2 * dms  = reinterpret_cast<const ggml_half2 *>(base + d_offset.second);
+
+        weights w;
+        w.bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
+
+        const int *      q4     = (const int *) (qs + 16 * w.bq8_offset + 4 * ((iqs / 2) % 4));
+        const uint16_t * scales = (const uint16_t *) scs;
+
+        w.v[0] = q4[0];
+        w.v[1] = q4[4];
+
+        const int j = (QR4_K * ((iqs / 2) / (QI8_1 / 2))) / 2;
+        if (j < 2) {
+            w.aux[0] = scales[j + 0] & 0x3f3f;
+            w.aux[1] = scales[j + 2] & 0x3f3f;
+        } else {
+            w.aux[0] = ((scales[j + 2] >> 0) & 0x0f0f) | ((scales[j - 2] & 0xc0c0) >> 2);
+            w.aux[1] = ((scales[j + 2] >> 4) & 0x0f0f) | ((scales[j - 0] & 0xc0c0) >> 2);
+        }
+
+        w.dm = *dms;
+
+        return w;
+    }
+
+    __dpct_inline__ static activations load_activations(const int8_t * q8_1_quant_ptr,
+                                                        const sycl::half2 * q8_1_ds, const int & iqs) {
+        activations a;
+        const int bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
+        for (int i = 0; i < QR4_K; ++i) {
+            const int8_t * quant_base_ptr = q8_1_quant_ptr + (bq8_offset + i) * QK8_1;
+            sycl::half2    ds_values      = *(q8_1_ds + bq8_offset + i);
+
+            a.d8[i] = ds_values[0];
+
+            const int * q8 = (const int *) quant_base_ptr + ((iqs / 2) % 4);
+            a.u[2 * i + 0] = q8[0];
+            a.u[2 * i + 1] = q8[4];
+        }
+
+        return a;
+    }
+
+    __dpct_inline__ static float apply(const weights & w, const activations & a) {
+        const uint8_t * sc = (const uint8_t *) w.aux;
+        const uint8_t * m  = sc + 2;
+
+        return vec_dot_q4_K_q8_1_impl_vmmq(w.v, a.u, sc, m, w.dm, a.d8);
+    }
+
+    __dpct_inline__ static float dot(const weights & w, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const auto a = load_activations(q8_1_quant_ptr, q8_1_ds, iqs);
+
+        return apply(w, a);
+    }
+
     __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
                                      const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
                                      const sycl::half2 * q8_1_ds, const int & iqs) {
-        const uint8_t *    base           = static_cast<const uint8_t *>(vbq);
-        const uint8_t *    qs             = base + ibx_offset.first;
-        const uint8_t *    scs            = base + d_offset.first;
-        const ggml_half2 * dms            = reinterpret_cast<const ggml_half2 *>(base + d_offset.second);
-
-        const int        bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
-        const int *      q4         = (const int *) (qs + 16 * bq8_offset + 4 * ((iqs / 2) % 4));
-        const uint16_t * scales     = (const uint16_t *) scs;
-
-        int   v[2];
-        int   u[2 * QR4_K];
-        float d8[QR4_K];
-
-        v[0] = q4[0];
-        v[1] = q4[4];
-
-        uint16_t  aux[2];
-        const int j = (QR4_K * ((iqs / 2) / (QI8_1 / 2))) / 2;
-        if (j < 2) {
-            aux[0] = scales[j + 0] & 0x3f3f;
-            aux[1] = scales[j + 2] & 0x3f3f;
-        } else {
-            aux[0] = ((scales[j + 2] >> 0) & 0x0f0f) | ((scales[j - 2] & 0xc0c0) >> 2);
-            aux[1] = ((scales[j + 2] >> 4) & 0x0f0f) | ((scales[j - 0] & 0xc0c0) >> 2);
-        }
-
-        const uint8_t * sc = (const uint8_t *) aux;
-        const uint8_t * m  = sc + 2;
-
-        for (int i = 0; i < QR4_K; ++i) {
-            const int8_t* quant_base_ptr = q8_1_quant_ptr + (bq8_offset + i) * QK8_1;
-            sycl::half2 ds_values = *(q8_1_ds + bq8_offset + i);
-
-            d8[i]                   = ds_values[0];
-
-            const int * q8 = (const int *) quant_base_ptr + ((iqs / 2) % 4);
-            u[2 * i + 0]   = q8[0];
-            u[2 * i + 1]   = q8[4];
-        }
-
-        return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, *dms, d8);
+        return dot(load(vbq, ibx_offset, d_offset, iqs), q8_1_quant_ptr, q8_1_ds, iqs);
     }
 };
 

ggml/src/ggml.c

@@ -7335,7 +7335,7 @@ void ggml_build_backward_expand(
         }
 
         // inplace operations are currently not supported
-        GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW ||
+        GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_SET_ROWS || node->op == GGML_OP_VIEW ||
             node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE);
 
         const size_t ihash = ggml_hash_find(&cgraph->visited_hash_set, node);

gguf-py/gguf/constants.py

@@ -697,6 +697,7 @@ class MODEL_TENSOR(IntEnum):
     FFN_DOWN_CHEXP       = auto()
     FFN_UP_CHEXP         = auto()
     FFN_EXP_PROBS_B      = auto()
+    FFN_EXP_PROBS_B_VL   = auto() # deepseek4 vision (bias for image tokens)
     FFN_GATE_TID2EID     = auto()
     MOE_LATENT_DOWN      = auto() # nemotron 3 super
     MOE_LATENT_UP        = auto() # nemotron 3 super
@@ -1449,6 +1450,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.FFN_UP_EXP:                "blk.{bid}.ffn_up_exps",
     MODEL_TENSOR.FFN_GATE_UP_EXP:           "blk.{bid}.ffn_gate_up_exps",
     MODEL_TENSOR.FFN_EXP_PROBS_B:           "blk.{bid}.exp_probs_b",
+    MODEL_TENSOR.FFN_EXP_PROBS_B_VL:        "blk.{bid}.exp_probs_b_vl",
     MODEL_TENSOR.FFN_GATE_TID2EID:          "blk.{bid}.ffn_gate_tid2eid",
     MODEL_TENSOR.MOE_LATENT_DOWN:           "blk.{bid}.ffn_latent_down",      # nemotron 3 super
     MODEL_TENSOR.MOE_LATENT_UP:             "blk.{bid}.ffn_latent_up",        # nemotron 3 super
@@ -3839,6 +3841,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_GATE_INP,
         MODEL_TENSOR.FFN_GATE_TID2EID,
         MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_EXP_PROBS_B_VL,
         MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_GATE_EXP,
         MODEL_TENSOR.FFN_DOWN_EXP,

gguf-py/gguf/gguf_writer.py

@@ -733,8 +733,11 @@ class GGUFWriter:
         else:
             self.add_array(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)
 
-    def add_expert_feed_forward_length(self, length: int) -> None:
-        self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+    def add_expert_feed_forward_length(self, length: int | Sequence[int]) -> None:
+        if isinstance(length, int):
+            self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+        else:
+            self.add_array(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
 
     def add_expert_shared_feed_forward_length(self, length: int) -> None:
         self.add_uint32(Keys.LLM.EXPERT_SHARED_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
@@ -860,8 +863,11 @@ class GGUFWriter:
     def add_expert_count(self, count: int) -> None:
         self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)
 
-    def add_expert_used_count(self, count: int) -> None:
-        self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count)
+    def add_expert_used_count(self, count: int | Sequence[int]) -> None:
+        if isinstance(count, int):
+            self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count)
+        else:
+            self.add_array(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count)
 
     def add_expert_shared_count(self, count: int) -> None:
         self.add_uint32(Keys.LLM.EXPERT_SHARED_COUNT.format(arch=self.arch), count)

src/llama-arch.cpp

@@ -457,6 +457,7 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_FFN_UP_SHEXP,                           "blk.%d.ffn_up_shexp" },
     { LLM_TENSOR_FFN_DOWN_SHEXP,                         "blk.%d.ffn_down_shexp" },
     { LLM_TENSOR_FFN_EXP_PROBS_B,                        "blk.%d.exp_probs_b" },
+    { LLM_TENSOR_FFN_EXP_PROBS_B_VL,                     "blk.%d.exp_probs_b_vl" },
     { LLM_TENSOR_FFN_LATENT_DOWN,                        "blk.%d.ffn_latent_down" },
     { LLM_TENSOR_FFN_LATENT_UP,                          "blk.%d.ffn_latent_up" },
     { LLM_TENSOR_ATTN_NORM_2,                            "blk.%d.attn_norm_2" },
@@ -896,6 +897,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_GATE_CHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_UP_CHEXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_EXP_PROBS_B,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_FFN_EXP_PROBS_B_VL,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     // altup / laurel (gemma 3n)
     {LLM_TENSOR_PER_LAYER_TOKEN_EMBD,       {LLM_TENSOR_LAYER_INPUT,     GGML_OP_GET_ROWS}},
     {LLM_TENSOR_PER_LAYER_MODEL_PROJ,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},

src/llama-arch.h

@@ -477,6 +477,7 @@ enum llm_tensor {
     LLM_TENSOR_FFN_GATE_CHEXPS,
     LLM_TENSOR_FFN_UP_CHEXPS,
     LLM_TENSOR_FFN_EXP_PROBS_B,
+    LLM_TENSOR_FFN_EXP_PROBS_B_VL,
     LLM_TENSOR_FFN_LATENT_DOWN,
     LLM_TENSOR_FFN_LATENT_UP,
     LLM_TENSOR_ATTN_Q_NORM,

src/llama-context.cpp

@@ -482,7 +482,8 @@ llama_context::~llama_context() {
     // wait for any pending asynchronous copies into the output buffers before they are freed
     synchronize();
 
-    if (!model.hparams.no_alloc) {
+    // when training, ggml_opt allocates extra buffers through the scheduler, so the sizes no longer match the expectation
+    if (!model.hparams.no_alloc && !opt_ctx) {
         for (size_t i = 0; i < backend_ptrs.size(); ++i) {
             ggml_backend_t             backend = backend_ptrs[i];
             ggml_backend_buffer_type_t buft    = backend_buft[i];
@@ -3408,6 +3409,15 @@ void llama_context::opt_init(struct llama_model * model, struct llama_opt_params
     GGML_ASSERT(model->hparams.n_ctx_train % n_batch  == 0);
     GGML_ASSERT(n_batch                    % n_ubatch == 0);
 
+    if (cparams.flash_attn) {
+        LLAMA_LOG_INFO("%s: disabling flash attention, FLASH_ATTN_EXT has no backward pass\n", __func__);
+        cparams.flash_attn = false;
+
+        // the graph changes without flash attention, need to reserve again
+        sched_need_reserve = true;
+        sched_reserve();
+    }
+
     ggml_opt_params opt_params = ggml_opt_default_params(sched.get(), GGML_OPT_LOSS_TYPE_CROSS_ENTROPY);
     opt_params.opt_period      = n_batch / n_ubatch;
     opt_params.get_opt_pars    = lopt_params.get_opt_pars;

src/llama-graph.cpp

@@ -1466,7 +1466,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     n_embd_head_v    (hparams.n_embd_head_v()),
     n_embd_v_gqa     (hparams.n_embd_v_gqa()),
     n_expert         (hparams.n_expert),
-    n_expert_used    (cparams.warmup ? hparams.n_expert : hparams.n_expert_used),
+    n_expert_used    (cparams.warmup ? hparams.n_expert : hparams.n_expert_used()),
     freq_base        (cparams.rope_freq_base),
     freq_scale       (cparams.rope_freq_scale),
     ext_factor       (cparams.yarn_ext_factor),
@@ -2270,25 +2270,26 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     assert(n_expert_used > 0);
 
     // order the views before the adds
-    for (uint32_t i = 0; i < hparams.n_expert_used; ++i) {
+    // Use per-layer n_expert_used to bound the graph even during warmup (avoids
+    // the large-add-nodes issue for uniform arches; for Puzzle the per-layer
+    // value is correct). ref: https://github.com/ggml-org/llama.cpp/pull/14753
+    const uint32_t n_expert_used_il = hparams.n_expert_used(il);
+    for (uint32_t i = 0; i < n_expert_used_il; ++i) {
         cur_experts[i] = ggml_view_2d(ctx0, experts, n_embd, n_tokens, experts->nb[2], i*experts->nb[1]);
 
         ggml_build_forward_expand(gf, cur_experts[i]);
     }
 
     // aggregate experts
-    // note: here we explicitly use hparams.n_expert_used instead of n_expert_used
-    //       to avoid potentially a large number of add nodes during warmup
-    //       ref: https://github.com/ggml-org/llama.cpp/pull/14753
     ggml_tensor * moe_out = cur_experts[0];
 
-    for (uint32_t i = 1; i < hparams.n_expert_used; ++i) {
+    for (uint32_t i = 1; i < n_expert_used_il; ++i) {
         moe_out = ggml_add(ctx0, moe_out, cur_experts[i]);
 
         ggml_build_forward_expand(gf, moe_out);
     }
 
-    if (hparams.n_expert_used == 1) {
+    if (n_expert_used_il == 1) {
         // avoid returning a non-contiguous tensor
         moe_out = ggml_cont(ctx0, moe_out);
     }

src/llama-hparams.cpp

@@ -71,6 +71,22 @@ uint32_t llama_hparams::n_ff(uint32_t il) const {
     GGML_ABORT("fatal error");
 }
 
+uint32_t llama_hparams::n_ff_exp(uint32_t il) const {
+    if (il < n_layer_all) {
+        return n_ff_exp_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_expert_used(uint32_t il) const {
+    if (il < n_layer_all) {
+        return n_expert_used_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
 uint32_t llama_hparams::n_gqa(uint32_t il) const {
     const uint32_t n_head    = this->n_head(il);
     const uint32_t n_head_kv = this->n_head_kv(il);

src/llama-hparams.h

@@ -62,7 +62,6 @@ struct llama_hparams {
     // per-token adapter selection. -1 when the model has no such layer.
     int32_t  router_layer = -1;
     uint32_t n_expert = 0;
-    uint32_t n_expert_used = 0;
     uint32_t n_rel_attn_bkts = 0;
 
     // TODO: this needs to be reworked
@@ -92,10 +91,14 @@ struct llama_hparams {
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
 
+    // per-layer expert feed-forward size
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_exp_arr;
+    // per-layer top-k expert routing count
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_expert_used_arr;
+
     uint32_t n_layer_dense_lead = 0;
     uint32_t n_lora_q           = 0;
     uint32_t n_lora_kv          = 0;
-    uint32_t n_ff_exp           = 0;
     uint32_t n_ff_shexp         = 0;
     uint32_t n_ff_chexp         = 0;
     uint32_t n_expert_shared    = 0;
@@ -161,6 +164,10 @@ struct llama_hparams {
     // the size of the sliding window (0 - no SWA)
     uint32_t n_swa = 0;
 
+    // deepseek4 vision: when decoding non-causally (multimodal input), SWA is not applied between tokens of the current ubatch (the image span); older tokens are still window-clipped
+    // for other models (like gemma 3, gemma 4): SWA is always applied to match transformers implementation
+    bool swa_full_non_causal = false;
+
     // if is_swa_impl[il] == 1, then layer il is SWA
     // if is_swa_impl[il] == 0, then layer il is dense (i.e. non-SWA)
     // by default, all layers are dense
@@ -381,6 +388,10 @@ struct llama_hparams {
 
     uint32_t n_ff(uint32_t il = 0) const;
 
+    uint32_t n_ff_exp(uint32_t il = 0) const;
+
+    uint32_t n_expert_used(uint32_t il = 0) const;
+
     uint32_t n_gqa(uint32_t il = 0) const;
 
     uint32_t n_rot(uint32_t il = 0) const;

src/llama-kv-cache.cpp

@@ -1681,7 +1681,9 @@ static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, T * data
 
                 // apply SWA if any
                 if (swa) {
-                    if (llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) {
+                    // see llama_hparams::swa_full_non_causal
+                    const bool in_span = !causal && args.hparams.swa_full_non_causal && p0 >= seq_pos_min[seq_id];
+                    if (!in_span && llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) {
                         goto skip;
                     }
                 }

src/llama-model-loader.cpp

@@ -951,7 +951,7 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
         case GGML_OP_MUL_MAT_ID:
             {
                 // Used for either MoE expert routing or embedded adapter routing
-                const int n_ids_used = hparams.router_layer >= 0 ? 1 : hparams.n_expert_used;
+                const int n_ids_used = hparams.router_layer >= 0 ? 1 : hparams.n_expert_used();
                 GGML_ASSERT(n_ids_used > 0);
                 ggml_tensor * b = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0], n_ids_used, 512);
                 ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_ids_used, 512);
@@ -964,7 +964,7 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
             } break;
         case GGML_OP_ADD_ID:
             {
-                const int n_expert_used = hparams.n_expert_used;
+                const int n_expert_used = hparams.n_expert_used();
                 GGML_ASSERT(n_expert_used > 0);
                 ggml_tensor * a = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0], n_expert_used, 512);
                 ggml_tensor * c = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_expert_used, 512);
@@ -1497,7 +1497,6 @@ bool llama_model_loader::load_all_data(
     }
     GGML_ASSERT(size_data != 0 && "call init_mappings() first");
 
-    std::vector<no_init<uint8_t>> read_buf;
     std::vector<std::future<std::pair<ggml_tensor *, bool>>> validation_result;
 
     // 4 staging buffers for async uploads, each sized 1MB seems to be a good default for single NVMe drives.
@@ -1598,7 +1597,25 @@ bool llama_model_loader::load_all_data(
             ggml_backend_name(upload_backend));
     }
 
+    std::vector<ggml_tensor *> tensors;
     for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) {
+        tensors.push_back(cur);
+    }
+
+    // without mmap, tensors in non-host buffers are staged through a temporary buffer sized like the tensor
+    // load them biggest-first so the largest staging buffer is allocated while the fewest weights are resident
+    if (!use_mmap) {
+        std::stable_sort(tensors.begin(), tensors.end(), [](const ggml_tensor * a, const ggml_tensor * b) {
+            const bool staged_a = a->buffer && !ggml_backend_buffer_is_host(a->buffer);
+            const bool staged_b = b->buffer && !ggml_backend_buffer_is_host(b->buffer);
+            if (staged_a != staged_b) {
+                return staged_a;
+            }
+            return staged_a && ggml_nbytes(a) > ggml_nbytes(b);
+        });
+    }
+
+    for (struct ggml_tensor * cur : tensors) {
         const auto * weight = get_weight(ggml_get_name(cur));
         if (weight == nullptr) {
             // this can happen with split experts models
@@ -1711,7 +1728,8 @@ bool llama_model_loader::load_all_data(
                         buffer_idx %= n_buffers;
                     }
                 } else {
-                    read_buf.resize(n_size);
+                    // scoped to one tensor so only one staging buffer is alive at a time
+                    std::vector<no_init<uint8_t>> read_buf(n_size);
                     file->seek(weight->offs, SEEK_SET);
                     file->read_raw(read_buf.data(), n_size);
                     ggml_backend_tensor_set(cur, read_buf.data(), 0, n_size);

src/llama-model-saver.cpp

@@ -222,7 +222,7 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_BLOCK_COUNT,                       hparams.n_layer_all);
     add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);
     add_kv(LLM_KV_FEED_FORWARD_LENGTH,               hparams.n_ff_arr, true);
-    add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp());
     add_kv(LLM_KV_EXPERT_LATENT_LENGTH,              hparams.n_expert_latent);
     add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp);
     add_kv(LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,  hparams.n_ff_chexp);
@@ -233,7 +233,7 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_USE_PARALLEL_RESIDUAL,             hparams.use_par_res);
     // add_kv(LLM_KV_TENSOR_DATA_LAYOUT,                ???);
     add_kv(LLM_KV_EXPERT_COUNT,                      hparams.n_expert);
-    add_kv(LLM_KV_EXPERT_USED_COUNT,                 hparams.n_expert_used);
+    add_kv(LLM_KV_EXPERT_USED_COUNT,                 hparams.n_expert_used());
     add_kv(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
     add_kv(LLM_KV_EXPERT_GROUP_COUNT,                hparams.n_expert_groups);
     add_kv(LLM_KV_EXPERT_GROUP_USED_COUNT,           hparams.n_group_used);

src/llama-model.cpp

@@ -634,7 +634,7 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
 
             // the FFN is the same for Qwen 3 Next and Qwen 3.5:
             if (std::regex_match(tensor_name, pattern_ffn_gate_up_weight)) {
-                const int64_t n_ff_exp = hparams.n_ff_exp;
+                const int64_t n_ff_exp = hparams.n_ff_exp(il);
                 GGML_ASSERT(tensor->ne[axis] == 2*n_ff_exp);
                 return {{n_ff_exp, 2}};
             }
@@ -657,7 +657,7 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
             return {{tensor->ne[axis], 1}};
         }
         if (std::regex_match(tensor_name, pattern_ffn_gate_up_weight)) {
-            const int64_t n_ff_exp = hparams.n_ff_exp;
+            const int64_t n_ff_exp = hparams.n_ff_exp(il);
             GGML_ASSERT(tensor->ne[axis] == 2*n_ff_exp);
             return {{n_ff_exp, 2}};
         }
@@ -943,6 +943,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
         case LLM_TYPE_35B_A3B:       return "35B.A3B";
         case LLM_TYPE_48B_A3B:       return "48B.A3B";
+        case LLM_TYPE_75B_A9B:       return "75B.A9B";
         case LLM_TYPE_80B_A3B:       return "80B.A3B";
         case LLM_TYPE_A3B:           return "A3B";
         case LLM_TYPE_100B_A6B:      return "100B.A6B";
@@ -1226,14 +1227,15 @@ void llama_model_base::load_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,    hparams.n_layer_nextn,   false);
     GGML_ASSERT(hparams.n_layer_nextn <= hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_COUNT,            hparams.n_expert,        false);
-    ml.get_key(LLM_KV_EXPERT_USED_COUNT,       hparams.n_expert_used,   false);
+    std::fill(hparams.n_expert_used_arr.begin(), hparams.n_expert_used_arr.end(), 0);
+    ml.get_key_or_arr(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_EXPERT_GROUP_COUNT,      hparams.n_expert_groups, false);
     ml.get_key(LLM_KV_EXPERT_GROUP_USED_COUNT, hparams.n_group_used,    false);
 
     if (arch == LLM_ARCH_HUNYUAN_VL || arch == LLM_ARCH_HUNYUAN_DENSE) {
         if (hparams.n_expert <= 1) {
-            hparams.n_expert      = 0;
-            hparams.n_expert_used = 0;
+            hparams.n_expert = 0;
+            std::fill(hparams.n_expert_used_arr.begin(), hparams.n_expert_used_arr.end(), 0);
         }
     }
 
@@ -1251,10 +1253,16 @@ void llama_model_base::load_hparams(llama_model_loader & ml) {
         GGML_ASSERT(hparams.convnext.n_layer <= hparams.n_layer_all);
     }
 
+    // models may route a different number of experts per layer, so validate the maximum
+    uint32_t n_expert_used_max = 0;
+    for (uint32_t il = 0; il < hparams.n_layer_all; ++il) {
+        n_expert_used_max = std::max(n_expert_used_max, hparams.n_expert_used(il));
+    }
+
     GGML_ASSERT(hparams.n_expert <= LLAMA_MAX_EXPERTS);
-    GGML_ASSERT(hparams.n_expert_used <= hparams.n_expert);
+    GGML_ASSERT(n_expert_used_max <= hparams.n_expert);
     if (hparams.n_expert > 0) {
-        GGML_ASSERT(hparams.n_expert_used > 0);
+        GGML_ASSERT(n_expert_used_max > 0);
         GGML_ASSERT(hparams.n_expert_groups < hparams.n_expert);
         if (hparams.n_expert_groups > 1) {
             GGML_ASSERT(hparams.n_expert % hparams.n_expert_groups == 0);
@@ -1262,13 +1270,14 @@ void llama_model_base::load_hparams(llama_model_loader & ml) {
             GGML_ASSERT(hparams.n_group_used < hparams.n_expert_groups);
         }
     } else {
-        GGML_ASSERT(hparams.n_expert_used == 0);
+        GGML_ASSERT(n_expert_used_max == 0);
         GGML_ASSERT(hparams.n_expert_groups == 0);
     }
 
-    std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
-    std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
-    std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
+    std::fill(hparams.n_head_arr.begin(),       hparams.n_head_arr.end(),       0);
+    std::fill(hparams.n_head_kv_arr.begin(),    hparams.n_head_kv_arr.end(),    0);
+    std::fill(hparams.n_ff_arr.begin(),         hparams.n_ff_arr.end(),         0);
+    std::fill(hparams.n_ff_exp_arr.begin(),     hparams.n_ff_exp_arr.end(),     0);
 
     std::fill(hparams.rope_sections.begin(), hparams.rope_sections.end(), 0);
     std::fill(hparams.rope_pattern.begin(),  hparams.rope_pattern.end(), 1);
@@ -1501,7 +1510,7 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
         const auto tn = LLM_TN(arch);
 
         const int64_t n_expert      = hparams.n_expert;
-        const int64_t n_expert_used = hparams.n_expert_used;
+        const int64_t n_expert_used = hparams.n_expert_used();
 
         if (n_expert > 0 && n_expert_used == 0) {
             throw std::runtime_error("model has expert layers but no expert layers are used");
@@ -1807,6 +1816,14 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
         return true;
     }
 
+    // without mmap, load non-host buffers first: their tensors go through a staging buffer, which is cheapest while the fewest weights are resident
+    if (!ml.use_mmap) {
+        std::stable_partition(ctx_buf_maps.begin(), ctx_buf_maps.end(), [](const auto & ctx_buf_map) {
+            const auto & buf_map = ctx_buf_map.second;
+            return !buf_map.empty() && !ggml_backend_buffer_is_host(buf_map.begin()->second);
+        });
+    }
+
     // load tensor data
     for (auto & [ctx, buf_map] : ctx_buf_maps) {
         if (!ml.load_all_data(ctx, buf_map, use_mlock ? &pimpl->mlock_mmaps : NULL, params.progress_callback, params.progress_callback_user_data)) {
@@ -1957,7 +1974,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: f_attn_value_scale    = %.4f\n",   __func__, hparams.f_attn_value_scale);
         LLAMA_LOG_INFO("%s: n_ff                  = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer_all).c_str());
         LLAMA_LOG_INFO("%s: n_expert              = %u\n",     __func__, hparams.n_expert);
-        LLAMA_LOG_INFO("%s: n_expert_used         = %u\n",     __func__, hparams.n_expert_used);
+        LLAMA_LOG_INFO("%s: n_expert_used         = %u\n",     __func__, hparams.n_expert_used());
         LLAMA_LOG_INFO("%s: n_expert_groups       = %d\n",     __func__, hparams.n_expert_groups);
         LLAMA_LOG_INFO("%s: n_group_used          = %d\n",     __func__, hparams.n_group_used);
         LLAMA_LOG_INFO("%s: causal attn           = %d\n",     __func__, hparams.causal_attn);
@@ -2032,7 +2049,7 @@ void llama_model::print_info() const {
 
         if (arch == LLM_ARCH_DEEPSEEK) {
             LLAMA_LOG_INFO("%s: n_layer_dense_lead    = %d\n",     __func__, hparams.n_layer_dense_lead);
-            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
+            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp());
             LLAMA_LOG_INFO("%s: n_expert_shared       = %d\n",     __func__, hparams.n_expert_shared);
             LLAMA_LOG_INFO("%s: expert_weights_scale  = %.1f\n",   __func__, hparams.expert_weights_scale);
         }
@@ -2045,7 +2062,7 @@ void llama_model::print_info() const {
             LLAMA_LOG_INFO("%s: n_lora_kv             = %d\n",     __func__, hparams.n_lora_kv);
             LLAMA_LOG_INFO("%s: n_embd_head_k_mla     = %d\n",     __func__, hparams.n_embd_head_k_mla());
             LLAMA_LOG_INFO("%s: n_embd_head_v_mla     = %d\n",     __func__, hparams.n_embd_head_v_mla());
-            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
+            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp());
             LLAMA_LOG_INFO("%s: n_expert_shared       = %d\n",     __func__, hparams.n_expert_shared);
             LLAMA_LOG_INFO("%s: expert_weights_scale  = %.1f\n",   __func__, hparams.expert_weights_scale);
             LLAMA_LOG_INFO("%s: expert_weights_norm   = %d\n",     __func__, hparams.expert_weights_norm);
@@ -2053,7 +2070,7 @@ void llama_model::print_info() const {
         }
 
         if (arch == LLM_ARCH_QWEN2MOE) {
-            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
+            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp());
             LLAMA_LOG_INFO("%s: n_ff_shexp            = %d\n",     __func__, hparams.n_ff_shexp);
         }
 
@@ -2063,7 +2080,7 @@ void llama_model::print_info() const {
                 arch == LLM_ARCH_OPENAI_MOE ||
                 arch == LLM_ARCH_QWEN3VLMOE ||
                 arch == LLM_ARCH_RND1) {
-            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
+            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp());
         }
 
         if (arch == LLM_ARCH_MINICPM ||
@@ -2080,7 +2097,7 @@ void llama_model::print_info() const {
 
         if (arch == LLM_ARCH_BAILINGMOE) {
             LLAMA_LOG_INFO("%s: n_layer_dense_lead    = %d\n",     __func__, hparams.n_layer_dense_lead);
-            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
+            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp());
             LLAMA_LOG_INFO("%s: n_expert_shared       = %d\n",     __func__, hparams.n_expert_shared);
             LLAMA_LOG_INFO("%s: expert_weights_scale  = %.1f\n",   __func__, hparams.expert_weights_scale);
             LLAMA_LOG_INFO("%s: expert_weights_norm   = %d\n",     __func__, hparams.expert_weights_norm);
@@ -2088,7 +2105,7 @@ void llama_model::print_info() const {
 
         if (arch == LLM_ARCH_BAILINGMOE2 || arch == LLM_ARCH_BAILINGMOE3) {
             LLAMA_LOG_INFO("%s: n_layer_dense_lead    = %d\n",     __func__, hparams.n_layer_dense_lead);
-            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
+            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp());
             LLAMA_LOG_INFO("%s: n_ff_shexp            = %d\n",     __func__, hparams.n_ff_shexp);
             LLAMA_LOG_INFO("%s: n_expert_shared       = %d\n",     __func__, hparams.n_expert_shared);
             LLAMA_LOG_INFO("%s: expert_weights_scale  = %.1f\n",   __func__, hparams.expert_weights_scale);
@@ -2098,12 +2115,12 @@ void llama_model::print_info() const {
         }
 
         if (arch == LLM_ARCH_SMALLTHINKER || arch == LLM_ARCH_LFM2MOE) {
-            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
+            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp());
             LLAMA_LOG_INFO("%s: expert_gating_func    = %s\n",     __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
         }
 
         if (arch == LLM_ARCH_GROVEMOE) {
-            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
+            LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp());
             LLAMA_LOG_INFO("%s: n_ff_chexp            = %d\n",     __func__, hparams.n_ff_chexp);
             LLAMA_LOG_INFO("%s: n_group_experts       = %d\n",     __func__, hparams.n_group_experts);
             LLAMA_LOG_INFO("%s: expert_group_scale    = %.2f\n",   __func__, hparams.expert_group_scale);

src/llama-model.h

@@ -128,6 +128,7 @@ enum llm_type {
     LLM_TYPE_31B_A3_5B,
     LLM_TYPE_35B_A3B, // Qwen3.5
     LLM_TYPE_48B_A3B, // Kimi Linear
+    LLM_TYPE_75B_A9B, // Nemotron 3 Puzzle
     LLM_TYPE_80B_A3B, // Qwen3 Next
     LLM_TYPE_A3B,     // Qwen3.8 Flash Next
     LLM_TYPE_100B_A6B,
@@ -362,6 +363,7 @@ struct llama_layer {
     struct ggml_tensor * ffn_up_b   = nullptr; // b3
     struct ggml_tensor * ffn_act    = nullptr;
     struct ggml_tensor * ffn_exp_probs_b = nullptr;
+    struct ggml_tensor * ffn_exp_probs_b_vl = nullptr; // deepseek4 vision (bias for image tokens)
     struct ggml_tensor * ffn_gate_tid2eid = nullptr;
 
     struct ggml_tensor * dflash_attn_conv_base = nullptr;
@@ -838,7 +840,7 @@ const char * llm_type_name(llm_type type);
     const int64_t n_token_types  = vocab.n_token_types();    GGML_UNUSED(n_token_types); \
     const int64_t n_rot          = hparams.n_rot();          GGML_UNUSED(n_rot); \
     const int64_t n_expert       = hparams.n_expert;         GGML_UNUSED(n_expert); \
-    const int64_t n_expert_used  = hparams.n_expert_used;    GGML_UNUSED(n_expert_used); \
+    const int64_t n_expert_used  = hparams.n_expert_used();  GGML_UNUSED(n_expert_used); \
     const int64_t n_ctx_train    = hparams.n_ctx_train;      GGML_UNUSED(n_ctx_train);
 
 // For internal test use

src/models/afmoe.cpp

@@ -3,7 +3,7 @@
 void llama_model_afmoe::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
@@ -52,7 +52,7 @@ void llama_model_afmoe::load_arch_tensors(llama_model_loader &) {
         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
     }
 
-    const int64_t n_ff_exp = hparams.n_ff_exp;
+    const int64_t n_ff_exp = hparams.n_ff_exp();
 
     for (int i = 0; i < n_layer; ++i) {
         auto & layer = layers[i];

src/models/bailingmoe.cpp

@@ -3,7 +3,7 @@
 void llama_model_bailingmoe::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
@@ -19,7 +19,7 @@ void llama_model_bailingmoe::load_arch_tensors(llama_model_loader &) {
     LLAMA_LOAD_LOCALS;
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
-    const int64_t n_ff_exp            = hparams.n_ff_exp;
+    const int64_t n_ff_exp            = hparams.n_ff_exp();
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 

src/models/bailingmoe2.cpp

@@ -3,7 +3,7 @@
 void llama_model_bailingmoe2::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
@@ -21,7 +21,7 @@ void llama_model_bailingmoe2::load_arch_tensors(llama_model_loader &) {
     LLAMA_LOAD_LOCALS;
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
-    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 

src/models/bailingmoe3.cpp

@@ -15,7 +15,7 @@ void llama_model_bailingmoe3::load_arch_hparams(llama_model_loader & ml) {
         hparams.kda_safe_gate = true;
     }
     ml.get_key(LLM_KV_KDA_GATE_LOWER_BOUND,             hparams.kda_gate_lower_bound);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,       hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,              hparams.n_expert_shared);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,        hparams.n_layer_dense_lead);
@@ -26,7 +26,7 @@ void llama_model_bailingmoe3::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_SHEXP,         hparams.swiglu_clamp_shexp, hparams.n_layer_all, false);
 
     if (hparams.n_ff_shexp == 0) {
-        hparams.n_ff_shexp = hparams.n_ff_exp * std::max(1u, hparams.n_expert_shared);
+        hparams.n_ff_shexp = hparams.n_ff_exp() * std::max(1u, hparams.n_expert_shared);
     }
 
     GGML_ASSERT(hparams.kda_safe_gate);
@@ -115,9 +115,9 @@ void llama_model_bailingmoe3::load_arch_tensors(llama_model_loader & ml) {
         } else {
             layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", il), { n_embd, n_expert }, trunk_flags);
             layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", il), { n_expert }, trunk_flags);
-            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", il), { n_embd, hparams.n_ff_exp, n_expert }, trunk_flags);
-            layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", il), { n_embd, hparams.n_ff_exp, n_expert }, trunk_flags);
-            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { hparams.n_ff_exp, n_embd, n_expert }, trunk_flags);
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", il), { n_embd, hparams.n_ff_exp(), n_expert }, trunk_flags);
+            layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", il), { n_embd, hparams.n_ff_exp(), n_expert }, trunk_flags);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { hparams.n_ff_exp(), n_embd, n_expert }, trunk_flags);
             layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", il), { n_embd, hparams.n_ff_shexp }, trunk_flags);
             layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", il), { n_embd, hparams.n_ff_shexp }, trunk_flags);
             layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", il), { hparams.n_ff_shexp, n_embd }, trunk_flags);
@@ -145,9 +145,9 @@ void llama_model_bailingmoe3::load_arch_tensors(llama_model_loader & ml) {
         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", il), { n_embd }, flags);
         layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", il), { n_embd, n_expert }, flags);
         layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", il), { n_expert }, flags);
-        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", il), { n_embd, hparams.n_ff_exp, n_expert }, flags);
-        layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", il), { n_embd, hparams.n_ff_exp, n_expert }, flags);
-        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { hparams.n_ff_exp, n_embd, n_expert }, flags);
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", il), { n_embd, hparams.n_ff_exp(), n_expert }, flags);
+        layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", il), { n_embd, hparams.n_ff_exp(), n_expert }, flags);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { hparams.n_ff_exp(), n_embd, n_expert }, flags);
         layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", il), { n_embd, hparams.n_ff_shexp }, flags);
         layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", il), { n_embd, hparams.n_ff_shexp }, flags);
         layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", il), { hparams.n_ff_shexp, n_embd }, flags);

src/models/bert.cpp

@@ -182,7 +182,7 @@ llama_model_bert::graph::graph(const llama_model & model, const llm_graph_params
                     nullptr,
                     model.layers[il].ffn_down_exps,
                     nullptr,
-                    hparams.n_expert, hparams.n_expert_used,
+                    hparams.n_expert, hparams.n_expert_used(),
                     LLM_FFN_GELU, false,
                     hparams.expert_weights_scale,
                     LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,

src/models/cohere2moe.cpp

@@ -13,7 +13,7 @@ void llama_model_cohere2moe::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
     ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
@@ -89,7 +89,7 @@ void llama_model_cohere2moe::load_arch_tensors(llama_model_loader & ml) {
             layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, flags);
             layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), { n_embd, n_ff }, flags);
         } else {
-            const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff;
+            const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff;
 
             layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert }, flags);
             layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, flags);
@@ -113,7 +113,7 @@ void llama_model_cohere2moe::load_arch_tensors(llama_model_loader & ml) {
         create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, flags);
         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags);
 
-        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff;
+        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff;
 
         // Routed experts
         layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert }, flags);

src/models/deepseek.cpp

@@ -3,11 +3,11 @@
 void llama_model_deepseek::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
 
-    switch (hparams.n_ff_exp) {
+    switch (hparams.n_ff_exp()) {
         case 1408: type = LLM_TYPE_16B; break;
         case 1792: type = LLM_TYPE_20B; break;
         default: type = LLM_TYPE_UNKNOWN;
@@ -19,7 +19,7 @@ void llama_model_deepseek::load_arch_tensors(llama_model_loader &) {
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
 
-    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 

src/models/deepseek2.cpp

@@ -15,7 +15,7 @@ void llama_model_deepseek2::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
     ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl, false);
     ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,        hparams.expert_weights_norm, false);
@@ -79,7 +79,7 @@ void llama_model_deepseek2::load_arch_tensors(llama_model_loader & ml) {
     const int64_t q_lora_rank  = hparams.n_lora_q;
     const int64_t kv_lora_rank = hparams.n_lora_kv;
 
-    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 

src/models/deepseek2ocr.cpp

@@ -4,7 +4,7 @@ void llama_model_deepseek2ocr::load_arch_hparams(llama_model_loader & ml) {
     // similar to deepseek2, but without MLA
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
@@ -25,7 +25,7 @@ void llama_model_deepseek2ocr::load_arch_tensors(llama_model_loader &) {
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
     // similar to deepseek2, but without MLA
-    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 

src/models/deepseek32.cpp

@@ -4,7 +4,7 @@
 #include "llama-kv-cache-dsa.h"
 
 void llama_model_deepseek32::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
     hparams.f_norm_eps = 1e-6;  // eps for layer norm
     ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
@@ -20,7 +20,7 @@ void llama_model_deepseek32::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
     ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl, false);
     ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
 
     // DSA parameters
@@ -71,7 +71,7 @@ void llama_model_deepseek32::load_arch_tensors(llama_model_loader & ml) {
     const int64_t q_lora_rank  = hparams.n_lora_q;
     const int64_t kv_lora_rank = hparams.n_lora_kv;
 
-    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);

src/models/deepseek4.cpp

@@ -29,7 +29,7 @@ void llama_model_deepseek4::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,       hparams.n_lora_q);
     ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
 
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm);
@@ -66,6 +66,9 @@ void llama_model_deepseek4::load_arch_hparams(llama_model_loader & ml) {
     }
     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
     hparams.set_swa_pattern(0);
+    // tokens of an image span attend bidirectionally to the whole span, the window only applies to older tokens
+    // ref: get_window_topk_idxs_visible in the reference impl
+    hparams.swa_full_non_causal = true;
     for (uint32_t il = hparams.n_layer(); il < hparams.n_layer_all; ++il) {
         hparams.is_swa_impl[il] = true;
     }
@@ -80,7 +83,7 @@ void llama_model_deepseek4::load_arch_tensors(llama_model_loader & ml) {
     LLAMA_LOAD_LOCALS;
 
     const int64_t q_lora_rank     = hparams.n_lora_q;
-    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
     const int64_t n_embd_head = hparams.n_embd_head_k();
@@ -156,6 +159,8 @@ void llama_model_deepseek4::load_arch_tensors(llama_model_loader & ml) {
         } else {
             layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, flags);
         }
+        // vision variant only: routing bias for image tokens
+        layer.ffn_exp_probs_b_vl = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B_VL, "bias", i), {n_expert}, flags | TENSOR_NOT_REQUIRED);
         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
 
         layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, flags);
@@ -1275,7 +1280,14 @@ llama_model_deepseek4::graph::graph(const llama_model & model, const llm_graph_p
         const auto & layer = model.layers[il];
         ggml_tensor * selected_experts = nullptr;
         ggml_tensor * exp_probs_b = layer.ffn_exp_probs_b;
-        if ((uint32_t) il < hparams.dsv4_hash_layer_count) {
+
+        // may apply exp_probs_b_vl is input is from mtmd
+        const bool is_media = ubatch.embd != nullptr;
+        if (is_media) {
+            if (layer.ffn_exp_probs_b_vl) {
+                exp_probs_b = layer.ffn_exp_probs_b_vl;
+            }
+        } else if ((uint32_t) il < hparams.dsv4_hash_layer_count) {
             selected_experts = ggml_get_rows(ctx0, layer.ffn_gate_tid2eid, res->t_inp_tokens);
             exp_probs_b = nullptr;
         }
@@ -1286,7 +1298,7 @@ llama_model_deepseek4::graph::graph(const llama_model & model, const llm_graph_p
                 layer.ffn_gate_exps,
                 layer.ffn_down_exps,
                 exp_probs_b,
-                n_expert, hparams.n_expert_used,
+                n_expert, hparams.n_expert_used(),
                 LLM_FFN_SILU, hparams.expert_weights_norm,
                 hparams.expert_weights_scale,
                 (llama_expert_gating_func_type) hparams.expert_gating_func,
@@ -1443,7 +1455,7 @@ llama_model_deepseek4::graph_mtp::graph_mtp(const llama_model & model, const llm
             layer.ffn_gate_exps,
             layer.ffn_down_exps,
             layer.ffn_exp_probs_b,
-            n_expert, hparams.n_expert_used,
+            n_expert, hparams.n_expert_used(),
             LLM_FFN_SILU, hparams.expert_weights_norm,
             hparams.expert_weights_scale,
             (llama_expert_gating_func_type) hparams.expert_gating_func,

src/models/dflash.cpp

@@ -40,7 +40,7 @@ void llama_model_dflash::load_arch_hparams(llama_model_loader & ml) {
     if (hparams.dsv4_hc_mult > 0) {
         ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,                hparams.n_lora_q);
         ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,             hparams.n_swa);
-        ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,           hparams.n_ff_exp);
+        ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,    hparams.n_ff_exp_arr, hparams.n_layer_all);
         ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,                  hparams.n_expert_shared);
         ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,                 hparams.expert_weights_scale);
         ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,                  hparams.expert_weights_norm);
@@ -159,7 +159,7 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
 
     if (hparams.dsv4_hc_mult > 0) {
         const int64_t q_lora_rank     = hparams.n_lora_q;
-        const int64_t n_ff_exp        = hparams.n_ff_exp;
+        const int64_t n_ff_exp        = hparams.n_ff_exp();
         const int64_t n_expert_shared = hparams.n_expert_shared;
         const int64_t n_embd_head     = hparams.n_embd_head_k();
         const int64_t o_groups        = hparams.dsv4_o_group_count;
@@ -948,7 +948,7 @@ llama_model_dflash::graph_dsv4::graph_dsv4(const llama_model & model, const llm_
                 layer.ffn_gate_exps,
                 layer.ffn_down_exps,
                 layer.ffn_exp_probs_b,
-                n_expert, hparams.n_expert_used,
+                n_expert, hparams.n_expert_used(),
                 LLM_FFN_SILU, hparams.expert_weights_norm,
                 hparams.expert_weights_scale,
                 (llama_expert_gating_func_type) hparams.expert_gating_func,

src/models/dots1.cpp

@@ -3,7 +3,7 @@
 void llama_model_dots1::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
@@ -19,7 +19,7 @@ void llama_model_dots1::load_arch_tensors(llama_model_loader &) {
     LLAMA_LOAD_LOCALS;
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
-    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 

src/models/dots3note.cpp

@@ -11,7 +11,7 @@ void llama_model_dots3note::load_arch_hparams(llama_model_loader & ml) {
 
     // MoE parameters
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,  hparams.n_layer_dense_lead);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,        hparams.expert_weights_norm, false);
@@ -56,7 +56,7 @@ void llama_model_dots3note::load_arch_tensors(llama_model_loader & ml) {
     const int64_t n_embd_head_qk_rope = hparams.n_rot();
 
     const int64_t q_lora_rank     = hparams.n_lora_q;
-    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);

src/models/ernie4-5.cpp

@@ -6,7 +6,7 @@ void llama_model_ernie4_5::load_arch_hparams(llama_model_loader & ml) {
 
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     if (arch == LLM_ARCH_ERNIE4_5_MOE) {
-        ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+        ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
         ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
         ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         hparams.n_moe_layer_step);
         ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);
@@ -47,7 +47,7 @@ void llama_model_ernie4_5::load_arch_tensors(llama_model_loader &) {
         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
         if (arch == LLM_ARCH_ERNIE4_5_MOE && static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) { // MoE layers
-            int n_ff_exp = hparams.n_ff_exp;
+            int n_ff_exp = hparams.n_ff_exp();
 
             layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
             layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);

src/models/exaone-moe.cpp

@@ -13,7 +13,7 @@ void llama_model_exaone_moe::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,          hparams.n_swa);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
@@ -30,7 +30,7 @@ void llama_model_exaone_moe::load_arch_hparams(llama_model_loader & ml) {
 void llama_model_exaone_moe::load_arch_tensors(llama_model_loader &) {
     LLAMA_LOAD_LOCALS;
 
-    const int64_t n_ff_exp       = hparams.n_ff_exp;
+    const int64_t n_ff_exp       = hparams.n_ff_exp();
     const int64_t n_ff_shexp     = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : n_ff_exp;
     const int64_t head_dim       = hparams.n_embd_head_k();
     const int64_t n_qo_dim       = n_head * head_dim;

src/models/gemma4.cpp

@@ -11,7 +11,7 @@ void llama_model_gemma4::load_arch_hparams(llama_model_loader & ml) {
     hparams.f_attention_scale     = 1.0f; // Gemma4 uses self.scaling = 1.0 (no pre-attn scaling)
 
     ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_EMBEDDING_LENGTH_PER_LAYER,  hparams.n_embd_per_layer);
@@ -32,7 +32,7 @@ void llama_model_gemma4::load_arch_tensors(llama_model_loader &) {
     LLAMA_LOAD_LOCALS;
 
     const uint32_t n_embd_per_layer = hparams.n_embd_per_layer;
-    const int64_t  n_ff_exp         = hparams.n_ff_exp;
+    const int64_t  n_ff_exp         = hparams.n_ff_exp();
 
     if (n_embd_head_k != n_embd_head_v) {
         throw std::runtime_error("Gemma 4 requires n_embd_head_k == n_embd_head_v");

src/models/glm-dsa.cpp

@@ -27,7 +27,7 @@ const std::array<uint32_t, LLAMA_MAX_LAYERS> GLM_5_2_DEFAULT_INDEXER_TYPES = {
 };
 
 void llama_model_glm_dsa::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
     ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
 
@@ -42,7 +42,7 @@ void llama_model_glm_dsa::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
     ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl, false);
     ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
 
     // DSA parameters
@@ -104,7 +104,7 @@ void llama_model_glm_dsa::load_arch_tensors(llama_model_loader & ml) {
     const int64_t q_lora_rank  = hparams.n_lora_q;
     const int64_t kv_lora_rank = hparams.n_lora_kv;
 
-    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 

src/models/glm4-moe.cpp

@@ -1,7 +1,7 @@
 #include "models.h"
 
 void llama_model_glm4_moe::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
     ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
 
@@ -40,7 +40,7 @@ void llama_model_glm4_moe::load_arch_tensors(llama_model_loader & ml) {
     }
 
     GGML_ASSERT(hparams.n_expert > 0 && "n_expert must be > 0 for GLM4_MOE MoE layers");
-    GGML_ASSERT(hparams.n_expert_used > 0 && "n_expert_used must be > 0 for GLM4_MOE MoE layers");
+    GGML_ASSERT(hparams.n_expert_used() > 0 && "n_expert_used must be > 0 for GLM4_MOE MoE layers");
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
 
@@ -82,7 +82,7 @@ void llama_model_glm4_moe::load_arch_tensors(llama_model_loader & ml) {
             layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), { n_expert }, flags);
 
             // MoE branch
-            const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+            const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
 
             layer.ffn_gate_exps = create_tensor(
                 tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags);

src/models/granite-swa.cpp

@@ -11,7 +11,7 @@ void llama_model_granite_swa::load_arch_hparams(llama_model_loader & ml) {
 
     // MoE expert configuration
     ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert, false);
-    ml.get_key(LLM_KV_EXPERT_USED_COUNT,           hparams.n_expert_used, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_USED_COUNT,    hparams.n_expert_used_arr, hparams.n_layer_all, false);
 
      // iSWA configuration
     ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);

src/models/grok.cpp

@@ -12,7 +12,7 @@ void llama_model_grok::load_arch_hparams(llama_model_loader & ml) {
     hparams.f_final_logit_softcapping = 0.0f;
 
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,  hparams.f_norm_rms_eps);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,   hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_LOGIT_SCALE,                  hparams.f_logit_scale, false);
     ml.get_key(LLM_KV_EMBEDDING_SCALE,              hparams.f_embedding_scale, false);
     ml.get_key(LLM_KV_ATTENTION_OUTPUT_SCALE,       hparams.f_attn_out_scale, false);
@@ -50,7 +50,7 @@ void llama_model_grok::load_arch_tensors(llama_model_loader &) {
         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
     }
 
-    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff/* / n_expert_used*/; // grok-1 n_ff_exp == n_ff
+    const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff/* / n_expert_used*/; // grok-1 n_ff_exp == n_ff
     for (int i = 0; i < n_layer; ++i) {
         auto & layer = layers[i];
 

src/models/grovemoe.cpp

@@ -1,7 +1,7 @@
 #include "models.h"
 
 void llama_model_grovemoe::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,  hparams.n_ff_chexp, false);
     ml.get_key(LLM_KV_EXPERT_GROUP_SCALE,                hparams.expert_group_scale);
     ml.get_key(LLM_KV_EXPERTS_PER_GROUP,                 hparams.n_group_experts);
@@ -46,7 +46,7 @@ void llama_model_grovemoe::load_arch_tensors(llama_model_loader &) {
         layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
 
         // MoE branch
-        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
         const int64_t n_ff_chexp = hparams.n_ff_chexp ? hparams.n_ff_chexp : n_embd_head_k;
         const int64_t n_chunk_expert = n_expert / hparams.n_group_experts;
 

src/models/hunyuan-moe.cpp

@@ -2,7 +2,7 @@
 
 void llama_model_hunyuan_moe::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
 
     switch (hparams.n_layer()) {

src/models/hy-v3.cpp

@@ -2,7 +2,7 @@
 
 void llama_model_hy_v3::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
@@ -45,7 +45,7 @@ void llama_model_hy_v3::load_arch_tensors(llama_model_loader & ml) {
 
     auto load_block = [&](int i, int flags) {
         auto & layer = layers[i];
-        const int64_t n_ff_exp   = hparams.n_ff_exp   ? hparams.n_ff_exp   : n_ff / (n_expert_used > 0 ? n_expert_used : 1);
+        const int64_t n_ff_exp   = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / (n_expert_used > 0 ? n_expert_used : 1);
         const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp;
 
         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);

src/models/kimi-k3.cpp

@@ -30,7 +30,7 @@ void llama_model_kimi_k3::load_arch_hparams(llama_model_loader & ml) {
         hparams.is_recr_impl[i] = hparams.n_head_kv(i) == 0;
     }
 
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,  hparams.n_layer_dense_lead, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
@@ -139,7 +139,7 @@ void llama_model_kimi_k3::load_arch_tensors(llama_model_loader &) {
             layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
             layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
         } else {
-            const int64_t n_ff_exp = hparams.n_ff_exp;
+            const int64_t n_ff_exp = hparams.n_ff_exp();
 
             layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, 0);
             layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias",   i), {n_expert}, 0);
@@ -584,7 +584,7 @@ ggml_tensor * llama_model_kimi_k3::graph::build_latent_moe(
         layer.ffn_down_exps,
         layer.ffn_exp_probs_b,
         hparams.n_expert,
-        hparams.n_expert_used,
+        hparams.n_expert_used(),
         LLM_FFN_SITU, hparams.expert_weights_norm,
         hparams.expert_weights_scale,
         (llama_expert_gating_func_type) hparams.expert_gating_func,

src/models/kimi-linear.cpp

@@ -19,7 +19,7 @@ void llama_model_kimi_linear::load_arch_hparams(llama_model_loader & ml) {
     }
 
     // MoE parameters - Kimi uses moe_intermediate_size = 1024
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
@@ -137,7 +137,7 @@ void llama_model_kimi_linear::load_arch_tensors(llama_model_loader &) {
         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
         // MoE intermediate size (different from dense FFN)
-        const int64_t n_ff_exp = hparams.n_ff_exp;
+        const int64_t n_ff_exp = hparams.n_ff_exp();
 
         // Kimi uses n_layer_dense_lead to determine which layers use dense FFN vs MoE
         // first_k_dense_replace = 1 means layer 0 uses dense FFN, layers 1+ use MoE
@@ -504,7 +504,7 @@ llama_model_kimi_linear::graph::graph(const llama_model & model, const llm_graph
                 layer.ffn_down_exps,
                 layer.ffn_exp_probs_b,
                 hparams.n_expert,
-                hparams.n_expert_used,
+                hparams.n_expert_used(),
                 LLM_FFN_SILU, true,
                 hparams.expert_weights_scale,
                 (llama_expert_gating_func_type) hparams.expert_gating_func,

src/models/laguna.cpp

@@ -9,7 +9,7 @@
 void llama_model_laguna::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
@@ -24,7 +24,7 @@ void llama_model_laguna::load_arch_hparams(llama_model_loader & ml) {
         // Weightless fixtures (test-llama-archs) omit this key; derive a nonzero
         // size so the shared expert is still built. Real GGUFs always carry the
         // exact value (routed and shared FF lengths may differ).
-        hparams.n_ff_shexp = hparams.n_ff_exp * hparams.n_expert_shared;
+        hparams.n_ff_shexp = hparams.n_ff_exp() * hparams.n_expert_shared;
     }
 
     // Sliding-window attention is OPTIONAL. XS.2 is hybrid (full / SWA / SWA /
@@ -76,7 +76,7 @@ void llama_model_laguna::load_arch_tensors(llama_model_loader & ml) {
         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
     }
 
-    const int64_t n_ff_exp   = hparams.n_ff_exp;
+    const int64_t n_ff_exp   = hparams.n_ff_exp();
     const int64_t n_ff_shexp = hparams.n_ff_shexp;
 
     for (int i = 0; i < n_layer; ++i) {

src/models/lfm2.cpp

@@ -53,9 +53,9 @@ void llama_model_lfm2::load_arch_tensors(llama_model_loader &) {
         if (is_moe_layer) {
             GGML_ASSERT(n_expert && n_expert_used);
             layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i),  {n_embd, n_expert}, 0);
-            layer.ffn_gate_exps   = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, hparams.n_ff_exp, n_expert}, 0);
-            layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {hparams.n_ff_exp,   n_embd, n_expert}, 0);
-            layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i),   {n_embd, hparams.n_ff_exp, n_expert}, 0);
+            layer.ffn_gate_exps   = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, hparams.n_ff_exp(), n_expert}, 0);
+            layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {hparams.n_ff_exp(),   n_embd, n_expert}, 0);
+            layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i),   {n_embd, hparams.n_ff_exp(), n_expert}, 0);
             layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
         } else {  // dense
             layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);

src/models/lfm2moe.cpp

@@ -6,7 +6,7 @@ void llama_model_lfm2moe::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_SHORTCONV_L_CACHE,           hparams.n_shortconv_l_cache);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func);
 
     for (uint32_t il = 0; il < hparams.n_layer(); ++il) {
@@ -42,9 +42,9 @@ void llama_model_lfm2moe::load_arch_tensors(llama_model_loader &) {
         if (is_moe_layer) {
             GGML_ASSERT(n_expert && n_expert_used);
             layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i),  {n_embd, n_expert}, 0);
-            layer.ffn_gate_exps   = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, hparams.n_ff_exp, n_expert}, 0);
-            layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {hparams.n_ff_exp,   n_embd, n_expert}, 0);
-            layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i),   {n_embd, hparams.n_ff_exp, n_expert}, 0);
+            layer.ffn_gate_exps   = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, hparams.n_ff_exp(), n_expert}, 0);
+            layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {hparams.n_ff_exp(),   n_embd, n_expert}, 0);
+            layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i),   {n_embd, hparams.n_ff_exp(), n_expert}, 0);
             layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
         } else {  // dense
             layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);

src/models/llada-moe.cpp

@@ -1,7 +1,7 @@
 #include "models.h"
 
 void llama_model_llada_moe::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
     // diffusion language model uses non-causal attention
@@ -39,7 +39,7 @@ void llama_model_llada_moe::load_arch_tensors(llama_model_loader &) {
 
         layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
 
-        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
 
         layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
         layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);

src/models/llama4.cpp

@@ -2,7 +2,7 @@
 
 void llama_model_llama4::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,   hparams.n_moe_layer_step);
 
     const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
@@ -75,7 +75,7 @@ void llama_model_llama4::load_arch_tensors(llama_model_loader &) {
         layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
 
         if (is_moe_layer) {
-            const int64_t n_ff_exp = hparams.n_ff_exp;
+            const int64_t n_ff_exp = hparams.n_ff_exp();
 
             layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
             layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);

src/models/mellum.cpp

@@ -2,7 +2,7 @@
 
 void llama_model_mellum::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
 
     if (hparams.n_swa > 0) {
@@ -61,7 +61,7 @@ void llama_model_mellum::load_arch_tensors(llama_model_loader &) {
             throw std::runtime_error("n_expert_used must be > 0 for Mellum");
         }
 
-        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
 
         layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
         layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);

src/models/mimo2.cpp

@@ -5,7 +5,7 @@ void llama_model_mimo2::load_arch_hparams(llama_model_loader & ml) {
 
     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
 
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,   hparams.n_swa);
     ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,         hparams.rope_freq_base_train_swa, false);
 
@@ -62,7 +62,7 @@ void llama_model_mimo2::load_arch_tensors(llama_model_loader & ml) {
         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED | flags);
 
         // MoE branch
-        int64_t n_ff_exp = hparams.n_ff_exp;
+        int64_t n_ff_exp = hparams.n_ff_exp();
         layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED | flags);
         layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp,   n_expert}, TENSOR_NOT_REQUIRED | flags);
         layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, TENSOR_NOT_REQUIRED | flags);

src/models/minimax-m2.cpp

@@ -2,7 +2,7 @@
 
 void llama_model_minimax_m2::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,  hparams.f_norm_rms_eps);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,   hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC,           hparams.expert_gating_func, false);
 
     switch (hparams.n_layer()) {

src/models/minimax-m3.cpp

@@ -13,7 +13,7 @@
 void llama_model_minimax_m3::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
@@ -36,7 +36,7 @@ void llama_model_minimax_m3::load_arch_hparams(llama_model_loader & ml) {
 void llama_model_minimax_m3::load_arch_tensors(llama_model_loader &) {
     LLAMA_LOAD_LOCALS;
     const int64_t n_expert_shared = hparams.n_expert_shared;
-    const int64_t n_ff_exp        = hparams.n_ff_exp;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 

src/models/nemotron-h.cpp

@@ -1,5 +1,7 @@
 #include "models.h"
 
+#include <algorithm> // std::max
+
 void llama_model_nemotron_h::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
     ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
@@ -16,7 +18,8 @@ void llama_model_nemotron_h::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,     hparams.f_norm_eps); // MTP head final_layernorm
 
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp,        false);
+    // Puzzle models set a different expert FFN size per layer
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp,      false);
     ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
@@ -26,7 +29,17 @@ void llama_model_nemotron_h::load_arch_hparams(llama_model_loader & ml) {
     switch (hparams.n_layer()) {
         case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B
         case 56: type = LLM_TYPE_9B; break;
-        case 88: type = LLM_TYPE_120B_A12B; break;
+        case 88:
+            {
+                // Nemotron 3 Super (uniform MoE) and Nemotron 3 Puzzle (per-layer
+                // heterogeneous MoE) both have 88 layers; the per-layer top-k array
+                // is the discriminator.
+                bool heterogeneous = false;
+                for (uint32_t i = 1; i < hparams.n_layer(); ++i) {
+                    heterogeneous |= hparams.n_expert_used_arr[i] != hparams.n_expert_used_arr[0];
+                }
+                type = heterogeneous ? LLM_TYPE_75B_A9B : LLM_TYPE_120B_A12B;
+            } break;
         default: type = LLM_TYPE_UNKNOWN;
     }
 }
@@ -94,7 +107,10 @@ void llama_model_nemotron_h::load_arch_tensors(llama_model_loader & ml) {
             layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
         }  else {
             if (n_expert != 0) {
-                const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+                // Use per-layer n_ff_exp; fall back to n_ff/n_expert_used if absent (existing GGUFs).
+                const int64_t n_ff_exp_i = hparams.n_ff_exp(i)
+                    ? (int64_t)hparams.n_ff_exp(i)
+                    : hparams.n_ff(i) / (int64_t)hparams.n_expert_used(i);
                 const int64_t n_ff_shexp = hparams.n_ff_shexp;
 
                 layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, trunk_flags);
@@ -104,8 +120,8 @@ void llama_model_nemotron_h::load_arch_tensors(llama_model_loader & ml) {
                 layer.ffn_latent_down = create_tensor(tn(LLM_TENSOR_FFN_LATENT_DOWN, "weight", i), {n_embd, moe_n_embd}, TENSOR_NOT_REQUIRED);
                 layer.ffn_latent_up   = create_tensor(tn(LLM_TENSOR_FFN_LATENT_UP,   "weight", i), {moe_n_embd, n_embd}, TENSOR_NOT_REQUIRED);
 
-                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   moe_n_embd, n_expert}, trunk_flags);
-                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {moe_n_embd, n_ff_exp, n_expert}, trunk_flags);
+                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp_i,   moe_n_embd, n_expert}, trunk_flags);
+                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {moe_n_embd, n_ff_exp_i, n_expert}, trunk_flags);
 
                 // Shared expert branch
                 layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, trunk_flags);
@@ -129,7 +145,7 @@ void llama_model_nemotron_h::load_arch_tensors(llama_model_loader & ml) {
         const int64_t n_head_i       = hparams.n_head(i);
         const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);
         const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
-        const int64_t n_ff_exp       = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_exp       = hparams.n_ff_exp(i) ? (int64_t)hparams.n_ff_exp(i) : n_ff / (int64_t)hparams.n_expert_used(i);
         const int64_t n_ff_shexp     = hparams.n_ff_shexp;
 
         // NextN input-fusion tensors
@@ -280,7 +296,7 @@ ggml_tensor * llama_model_nemotron_h::graph::build_ffn_layer(ggml_tensor * cur,
                     nullptr, // no gate
                     model.layers[il].ffn_down_exps,
                     model.layers[il].ffn_exp_probs_b,
-                    n_expert, n_expert_used,
+                    n_expert, (int64_t)hparams.n_expert_used(il),
                     LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
                     hparams.expert_weights_scale,
                     LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,

src/models/openai-moe.cpp

@@ -2,7 +2,7 @@
 
 void llama_model_openai_moe::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
 
     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
@@ -24,7 +24,7 @@ void llama_model_openai_moe::load_arch_hparams(llama_model_loader & ml) {
 void llama_model_openai_moe::load_arch_tensors(llama_model_loader &) {
     LLAMA_LOAD_LOCALS;
 
-    const int64_t n_ff_exp = hparams.n_ff_exp;
+    const int64_t n_ff_exp = hparams.n_ff_exp();
 
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 

src/models/qwen2moe.cpp

@@ -1,7 +1,7 @@
 #include "models.h"
 
 void llama_model_qwen2moe::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
 
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -42,7 +42,7 @@ void llama_model_qwen2moe::load_arch_tensors(llama_model_loader &) {
         }
 
         // MoE branch
-        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
 
         layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
         layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);

src/models/qwen35moe.cpp

@@ -2,7 +2,7 @@
 #include "llama-memory-recurrent.h"
 
 void llama_model_qwen35moe::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
 
@@ -54,7 +54,7 @@ void llama_model_qwen35moe::load_arch_tensors(llama_model_loader & ml) {
     auto load_block_trunk = [&](int il, int flags) {
         auto & layer = layers[il];
 
-        const int64_t n_ff_exp   = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_exp   = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
         const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
 
         // Calculate dimensions from hyperparameters
@@ -106,7 +106,7 @@ void llama_model_qwen35moe::load_arch_tensors(llama_model_loader & ml) {
     auto load_block_mtp = [&](int il) {
         auto & layer = layers[il];
 
-        const int64_t n_ff_exp   = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_exp   = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
         const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
 
         // MTP block looks like a full-attention Qwen3.5 decoder block with MoE FFN.

src/models/qwen3moe.cpp

@@ -1,7 +1,7 @@
 #include "models.h"
 
 void llama_model_qwen3moe::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
     switch (hparams.n_layer()) {
@@ -47,7 +47,7 @@ void llama_model_qwen3moe::load_arch_tensors(llama_model_loader &) {
         }
 
         // MoE branch
-        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
 
         layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
         layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);

src/models/qwen3next.cpp

@@ -2,7 +2,7 @@
 #include "llama-memory-recurrent.h"
 
 void llama_model_qwen3next::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
 
@@ -50,7 +50,7 @@ void llama_model_qwen3next::load_arch_tensors(llama_model_loader & ml) {
         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
     }
 
-    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+    const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
 
     // Calculate dimensions from hyperparameters
     const int64_t head_k_dim = hparams.ssm_d_state;

src/models/qwen3vlmoe.cpp

@@ -3,7 +3,7 @@
 void llama_model_qwen3vlmoe::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_NUM_DEEPSTACK_LAYERS, hparams.n_deepstack_layers, false);
     ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
     switch (hparams.n_layer()) {
@@ -49,7 +49,7 @@ void llama_model_qwen3vlmoe::load_arch_tensors(llama_model_loader &) {
         }
 
         // MoE branch
-        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
 
         layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
         layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);

src/models/qwen4exp.cpp

@@ -24,7 +24,7 @@ static void qwen4exp_require_arr_len(llama_model_loader & ml, llm_kv kid, uint32
 }
 
 void llama_model_qwen4exp::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
 
@@ -191,7 +191,7 @@ void llama_model_qwen4exp::load_arch_tensors(llama_model_loader & ml) {
     for (int il = 0; il < n_layer; ++il) {
         auto & layer = layers[il];
 
-        const int64_t n_ff_exp   = hparams.n_ff_exp   ? hparams.n_ff_exp   : n_ff / n_expert_used;
+        const int64_t n_ff_exp   = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
         const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
 
         const int64_t head_k_dim = hparams.ssm_d_state;

src/models/rnd1.cpp

@@ -1,7 +1,7 @@
 #include "models.h"
 
 void llama_model_rnd1::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
     switch (hparams.n_layer()) {
@@ -49,7 +49,7 @@ void llama_model_rnd1::load_arch_tensors(llama_model_loader &) {
         }
 
         // MoE branch
-        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
 
         layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
         layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);

src/models/smallthinker.cpp

@@ -18,7 +18,7 @@ void llama_model_smallthinker::load_arch_hparams(llama_model_loader & ml) {
         hparams.n_no_rope_layer_step = hparams.n_layer();
     }
 
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp, false);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
 
@@ -57,7 +57,7 @@ void llama_model_smallthinker::load_arch_tensors(llama_model_loader &) {
         GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for SMALLTHINKER");
 
         // MoE branch
-        const int64_t n_ff_exp = hparams.n_ff_exp;
+        const int64_t n_ff_exp = hparams.n_ff_exp();
         layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0);
         layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
         layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0);

src/models/step35.cpp

@@ -9,7 +9,7 @@ void llama_model_step35::load_arch_hparams(llama_model_loader & ml) {
     hparams.n_rot_full = hparams.n_rot_full / 2;
 
     // MoE + SWA parameters
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
     ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func, false);
     ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
@@ -99,7 +99,7 @@ void llama_model_step35::load_arch_tensors(llama_model_loader & ml) {
         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED);
 
         // MoE routed experts + selection bias (router_bias)
-        const int64_t n_ff_exp = hparams.n_ff_exp;
+        const int64_t n_ff_exp = hparams.n_ff_exp();
         layer.ffn_gate_inp      = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);
         layer.ffn_gate_exps     = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp,   n_expert}, TENSOR_NOT_REQUIRED);
         layer.ffn_down_exps     = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, TENSOR_NOT_REQUIRED);
@@ -150,7 +150,7 @@ void llama_model_step35::load_arch_tensors(llama_model_loader & ml) {
         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED);
 
         // MoE routed experts + selection bias (router_bias)
-        const int64_t n_ff_exp = hparams.n_ff_exp;
+        const int64_t n_ff_exp = hparams.n_ff_exp();
         layer.ffn_gate_inp      = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);
         layer.ffn_gate_exps     = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp,   n_expert}, TENSOR_NOT_REQUIRED);
         layer.ffn_down_exps     = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, TENSOR_NOT_REQUIRED);

tests/test-backend-ops.cpp

@@ -9435,6 +9435,21 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_mul_mat(type_a,    GGML_TYPE_F32, 16,  8, 16*256, { 1,  1}, {1, 1}));
     }
 
+    // Multi-column MMVQ coverage for the Q4_K weight-reuse path and a Q5_K control.
+    for (ggml_type type_a : { GGML_TYPE_Q4_K, GGML_TYPE_Q5_K }) {
+        for (int n = 1; n <= 8; ++n) {
+            test_cases.emplace_back(new test_mul_mat(type_a, GGML_TYPE_F32, 4096, n, 1024, { 1, 1 }, { 1, 1 }));
+            test_cases.emplace_back(new test_mul_mat(type_a, GGML_TYPE_F32, 1023, n, 4096, { 1, 1 }, { 1, 1 }));
+        }
+    }
+
+    // The SYCL backend picks between one and two output rows per subgroup by row count when there
+    // are two destination columns (Q4_K_MMVQ_ROW_PAIR_MIN_NROWS in ggml-sycl/mmvq.cpp). Cover both
+    // sides of that boundary, including an odd row count above it for the row-pair tail.
+    for (int64_t m : {6271, 6272, 6273}) {
+        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_K, GGML_TYPE_F32, m, 2, 1024, { 1, 1 }, { 1, 1 }));
+    }
+
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 2880, 32, 2880, {1, 1}, {1, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 2880, 32, 2880, {1, 1}, {1, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_MXFP4, GGML_TYPE_F32, 2880, 32, 2880, {1, 1}, {1, 1}));
@@ -10364,6 +10379,22 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             true, 16, 8, b, false, true, false));
     }
 
+    // Fused row-pair coverage: minimum rows, an even pair, and an odd tail.
+    for (ggml_glu_op glu_op : { GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU }) {
+        for (int64_t m_batch : { 2, 3, 4 }) {
+            for (int64_t rows : { 1, 2, 3 }) {
+                test_cases.emplace_back(new test_mul_mat_vec_fusion(GGML_TYPE_Q4_K, glu_op, m_batch, rows, 256,
+                    false, 16, 8, false, false, true, false, { 1, 1 }));
+            }
+        }
+    }
+
+    // Both sides of the same row-count boundary as above, on the fused path.
+    for (int64_t rows : {6271, 6272, 6273}) {
+        test_cases.emplace_back(new test_mul_mat_vec_fusion(GGML_TYPE_Q4_K, GGML_GLU_OP_SWIGLU, 2, rows, 256,
+            false, 16, 8, false, false, true, false, { 1, 1 }));
+    }
+
     for (auto gate : {GATING_FUNC_SOFTMAX, GATING_FUNC_SIGMOID, GATING_FUNC_SOFTMAX_WEIGHT, GATING_FUNC_SQRT_SOFTPLUS}) {
         for (bool with_norm : {false, true}) {
             for (bool bias_probs : {false, true}) {
@@ -10651,6 +10682,16 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
         }
     }
 
+    // Q4_K multi-column mat-vec, at ffn_up/ffn_gate geometry (k = n_embd, m = n_ff): n sweeps the
+    // per-column specializations used for short prompts and speculative/MTP verify, and m brackets
+    // the row count at which the SYCL backend switches to two output rows per subgroup
+    // (Q4_K_MMVQ_ROW_PAIR_MIN_NROWS in ggml-sycl/mmvq.cpp), so both sides of it can be measured.
+    for (int64_t m : {4096, 6144, 6272, 14336}) {
+        for (int bs : {1, 2, 3, 4, 8}) {
+            test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_K, GGML_TYPE_F32, m, bs, 4096, {1, 1}, {1, 1}));
+        }
+    }
+
     // qwen3-30b-a3b
     for (int bs : {1, 4, 8, 32, 64, 128, 256, 512}) {
         for (ggml_type type_a : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0, GGML_TYPE_Q4_K, GGML_TYPE_Q6_K, GGML_TYPE_IQ2_XS}) {

tests/test-mtmd-c-api.c

@@ -130,6 +130,39 @@ int main(void) {
     }
     printf("Chunk save/load round-trip OK\n");
 
+    // test input validation of mtmd_tokenize_from_parts()
+    // invalid parts are rejected before the ctx is used, so NULL ctx is OK here
+    {
+        mtmd_input_chunks * out = mtmd_input_chunks_init();
+        mtmd_bitmap * bmp = mtmd_bitmap_init(4, 4, NULL); // placeholder bitmap
+        struct mtmd_input_text txt = { "hello", 5, false, false };
+        struct mtmd_input_text txt_null = { NULL, 0, false, false };
+
+        struct mtmd_input_part part_both      = { &txt, bmp };
+        struct mtmd_input_part part_neither   = { NULL, NULL };
+        struct mtmd_input_part part_null_text = { &txt_null, NULL };
+        const mtmd_input_part * parts[1];
+        int32_t rc;
+
+        parts[0] = &part_both;
+        rc = mtmd_tokenize_from_parts(NULL, out, parts, 1, false);
+        printf("tokenize part with both text and bitmap rc = %d (expect 1)\n", rc);
+        assert(rc == 1);
+
+        parts[0] = &part_neither;
+        rc = mtmd_tokenize_from_parts(NULL, out, parts, 1, false);
+        printf("tokenize part with neither text nor bitmap rc = %d (expect 1)\n", rc);
+        assert(rc == 1);
+
+        parts[0] = &part_null_text;
+        rc = mtmd_tokenize_from_parts(NULL, out, parts, 1, false);
+        printf("tokenize part with null text pointer rc = %d (expect 1)\n", rc);
+        assert(rc == 1);
+
+        mtmd_bitmap_free(bmp);
+        mtmd_input_chunks_free(out);
+    }
+
     // Free the chunks
     mtmd_input_chunks_free(chunks);
 

tests/test-mtmd-impl.cpp

@@ -80,7 +80,7 @@ MAKE_TEST(test_temporal_merge_grouping) {
     // spec chars:
     //   v = video frame, w = video frame of another size, a = audio, i = plain image, t = text
     auto make_parts = [&pool](const std::string & spec) {
-        std::vector<mtmd_input_part> parts;
+        std::vector<mtmd_internal_part> parts;
         for (char c : spec) {
             if (c == 't') {
                 parts.push_back({ "hello", nullptr });

tools/mtmd/README-dev.md

@@ -20,6 +20,7 @@ In short:
 A typical pipeline of the core libmtmd is as follows:
 - A bitmap (RGB image or PCM audio) is created
 - Bitmap and the text prompt is provided to `mtmd_tokenize()` that breaks the input into chunks
+    - Alternatively, `mtmd_tokenize_from_parts()` takes a list of pre-split text/media parts instead of a marker-based prompt
     - The tokenizer function first expands a "lazy" bitmap if it finds one. Typically, this is used by video, so that one media token corresponds to one input bitmap
     - For models that support "fused" temporal frames like Qwen-VL, the tokenizer tries to merge pair of consecutive frames into one batch. Only bitmaps marked by `mtmd_bitmap_set_mergeable()` are merged
     - The preprocessor will then be called, which produces a list of chunks

tools/mtmd/mtmd-cli.cpp

@@ -109,16 +109,15 @@ struct mtmd_cli_context {
     mtmd_cli_context(common_params & params) : llama_init(common_init_from_params(params)) {
         model = llama_init->model();
         lctx = llama_init->context();
+        if (!model || !lctx) {
+            exit(1);
+        }
         vocab = llama_model_get_vocab(model);
         smpl = common_sampler_init(model, params.sampling);
         n_threads = params.cpuparams.n_threads;
         batch = llama_batch_init(1, 0, 1); // batch for next token generation
         n_batch = params.n_batch;
 
-        if (!model || !lctx) {
-            exit(1);
-        }
-
         init_vision_context(params);
 
         if (!mtmd_helper_model_can_chat(lctx, ctx_vision.get())) {
@@ -265,21 +264,50 @@ static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg) {
     auto formatted_chat = chat_add_and_format(ctx, msg);
     LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.c_str());
 
-    mtmd_input_text text;
-    text.text          = formatted_chat.data();
-    text.text_len      = formatted_chat.size();
-    text.add_special   = add_bos;
-    text.parse_special = true;
-
     if (g_is_interrupted) return 0;
 
-    mtmd::input_chunks chunks(mtmd_input_chunks_init());
+    // note: we replace the marker here instead of letting mtmd_tokenize() to do that
+    //       because we want to demonstrate how to use mtmd_tokenize_from_parts()
+
+    // split the formatted chat on the media marker to get text segments
+    const std::string marker = mtmd_default_marker();
+    std::vector<std::string> segments;
+    size_t start = 0;
+    size_t pos;
+    while ((pos = formatted_chat.find(marker, start)) != std::string::npos) {
+        segments.push_back(formatted_chat.substr(start, pos - start));
+        start = pos + marker.size();
+    }
+    segments.push_back(formatted_chat.substr(start));
+
     auto bitmaps_c_ptr = ctx.bitmaps.c_ptr();
-    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(),
+    if (segments.size() - 1 != bitmaps_c_ptr.size()) {
+        LOG_ERR("Number of media markers (%zu) does not match number of loaded media (%zu)\n",
+                segments.size() - 1, bitmaps_c_ptr.size());
+        return 1;
+    }
+
+    // interleave text and media parts
+    std::vector<mtmd_input_text> texts(segments.size());
+    std::vector<mtmd_input_part> parts;
+    for (size_t i = 0; i < segments.size(); i++) {
+        texts[i] = {segments[i].data(), segments[i].size(), /* add_special */ false, /* parse_special */ true};
+        parts.push_back({&texts[i], nullptr});
+        if (i < bitmaps_c_ptr.size()) {
+            parts.push_back({nullptr, bitmaps_c_ptr[i]});
+        }
+    }
+    std::vector<const mtmd_input_part *> parts_ptr;
+    for (const auto & p : parts) {
+        parts_ptr.push_back(&p);
+    }
+
+    mtmd::input_chunks chunks(mtmd_input_chunks_init());
+    int32_t res = mtmd_tokenize_from_parts(ctx.ctx_vision.get(),
                         chunks.ptr.get(), // output
-                        &text, // text
-                        bitmaps_c_ptr.data(),
-                        bitmaps_c_ptr.size());
+                        parts_ptr.data(),
+                        parts_ptr.size(),
+                        add_bos);
     if (res != 0) {
         LOG_ERR("Unable to tokenize prompt, res = %d\n", res);
         return 1;

tools/mtmd/mtmd-image.cpp

@@ -980,6 +980,56 @@ mtmd_image_preproc_out mtmd_image_preprocessor_idefics3::preprocess(const clip_i
     //
     // CITE: https://github.com/huggingface/transformers/blob/main/src/transformers/models/idefics3/image_processing_idefics3.py#L737
     const clip_image_size original_size = img.get_size();
+
+    // old gguf files have no preprocessor longest size, custom token limits also need the generic size below
+    if (hparams.image_longest_edge > 0 && hparams.image_min_pixels <= 0 && hparams.image_max_pixels <= 0) {
+        const int    tile_size    = hparams.image_size;
+        const int    longest_edge = hparams.image_longest_edge;
+        const double aspect_ratio = (double) original_size.width / original_size.height;
+
+        clip_image_size resized_size;
+        if (original_size.width >= original_size.height) {
+            resized_size.width   = longest_edge;
+            resized_size.height  = (int) (longest_edge / aspect_ratio);
+            resized_size.height += resized_size.height % 2;
+        } else {
+            resized_size.height  = longest_edge;
+            resized_size.width   = (int) (longest_edge * aspect_ratio);
+            resized_size.width  += resized_size.width % 2;
+        }
+
+        const int grid_x = (resized_size.width  + tile_size - 1) / tile_size;
+        const int grid_y = (resized_size.height + tile_size - 1) / tile_size;
+        const clip_image_size refined_size = clip_image_size{grid_x * tile_size, grid_y * tile_size};
+
+        clip_image_u8 resized_img;
+        img_tool::resize(img, resized_img, resized_size, hparams.image_resize_algo, PAD_NONE);
+
+        clip_image_u8 refined_img;
+        img_tool::resize(resized_img, refined_img, refined_size, hparams.image_resize_algo, PAD_NONE);
+
+        clip_image_u8 overview;
+        img_tool::resize(refined_img, overview, {tile_size, tile_size}, hparams.image_resize_algo, PAD_NONE);
+
+        std::vector<clip_image_u8> slices;
+        for (int y = 0; y < grid_y; y++) {
+            for (int x = 0; x < grid_x; x++) {
+                clip_image_u8 slice;
+                img_tool::crop(refined_img, slice, x * tile_size, y * tile_size, tile_size, tile_size);
+                slices.push_back(std::move(slice));
+            }
+        }
+
+        LOG_DBG("%s: grid size: %d x %d (%d tiles) + overview\n", __func__, grid_x, grid_y, grid_x * grid_y);
+
+        mtmd_image_preproc_out output;
+        output.append_overview(hparams, overview, true);
+        output.append(hparams, slices, true);
+        output.grid_x = grid_x;
+        output.grid_y = grid_y;
+        return output;
+    }
+
     const clip_image_size refined_size = img_tool::calc_size_preserved_ratio(
         original_size,
         { hparams.image_size, std::max(0, hparams.image_min_pixels), std::max(0, hparams.image_max_pixels), hparams.image_longest_edge });

tools/mtmd/mtmd-internal.h

@@ -10,10 +10,12 @@
 #define MTMD_INTERNAL_HEADER
 
 // bitmap is null for text parts
-struct mtmd_input_part {
+struct mtmd_internal_part {
     std::string text;
     const mtmd_bitmap * bitmap;
+    // only used for text parts
+    bool parse_special = false;
 };
 
 // [QWEN_VIDEO] merged parts are erased from `parts`, so one group always maps to one part
-std::vector<std::vector<const mtmd_bitmap *>> mtmd_group_mergeable_bitmaps(std::vector<mtmd_input_part> & parts, int n_merge);
+std::vector<std::vector<const mtmd_bitmap *>> mtmd_group_mergeable_bitmaps(std::vector<mtmd_internal_part> & parts, int n_merge);

tools/mtmd/mtmd.cpp

@@ -1097,7 +1097,7 @@ void mtmd_free(mtmd_context * ctx) {
     delete ctx;
 }
 
-std::vector<std::vector<const mtmd_bitmap *>> mtmd_group_mergeable_bitmaps(std::vector<mtmd_input_part> & parts, int n_merge) {
+std::vector<std::vector<const mtmd_bitmap *>> mtmd_group_mergeable_bitmaps(std::vector<mtmd_internal_part> & parts, int n_merge) {
     std::vector<std::vector<const mtmd_bitmap *>> output;
     for (size_t i = 0; i < parts.size(); i++) {
         if (parts[i].bitmap == nullptr) {
@@ -1124,7 +1124,7 @@ struct mtmd_tokenizer {
     bool parse_special;
     const llama_vocab * vocab;
 
-    using part = mtmd_input_part;
+    using part = mtmd_internal_part;
     std::vector<part> parts;
     // these will be freed when mtmd_tokenizer finishes
     std::vector<mtmd::bitmap> bm_from_lazy; // TODO @ngxson : refactor, free bm_from_lazy progressively
@@ -1160,7 +1160,7 @@ struct mtmd_tokenizer {
                 }
                 parts.push_back({"", bitmaps[i_bm++]});
             } else {
-                parts.push_back({std::move(part), nullptr});
+                parts.push_back({std::move(part), nullptr, parse_special});
             }
         }
 
@@ -1177,6 +1177,26 @@ struct mtmd_tokenizer {
         expand_lazy_bitmaps();
     }
 
+    mtmd_tokenizer(mtmd_context * ctx,
+            const mtmd_input_part ** input_parts,
+            size_t n_parts,
+            bool add_special) : ctx(ctx) {
+        this->add_special = add_special;
+        parse_special = true; // only used for text returned by lazy bitmaps
+        vocab         = ctx->vocab;
+
+        for (size_t i = 0; i < n_parts; i++) {
+            const mtmd_input_part * p = input_parts[i];
+            if (p->text != nullptr) {
+                parts.push_back({std::string(p->text->text, p->text->text_len), nullptr, p->text->parse_special});
+            } else {
+                parts.push_back({"", p->bitmap});
+            }
+        }
+
+        expand_lazy_bitmaps();
+    }
+
     void expand_lazy_bitmaps() {
         std::vector<part> expanded;
         expanded.reserve(parts.size());
@@ -1201,7 +1221,7 @@ struct mtmd_tokenizer {
                             LOG_DBG("%s: lazy callback returned bitmap with dimensions %d x %d\n", __func__, out_bm->nx, out_bm->ny);
                         } else if (out_str) {
                             auto & ptr = text_from_lazy.emplace_back(out_str); // remember to free it later
-                            expanded.push_back({ptr, nullptr});
+                            expanded.push_back({ptr, nullptr, parse_special});
                             LOG_DBG("%s: lazy callback returned text: %s\n", __func__, out_str);
                         }
                     } else if (res == -1) {
@@ -1245,7 +1265,7 @@ struct mtmd_tokenizer {
                     return res;
                 }
             } else {
-                add_text(p.text, parse_special);
+                add_text(p.text, p.parse_special);
             }
         }
 
@@ -1727,6 +1747,30 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
     }
 }
 
+int32_t mtmd_tokenize_from_parts(mtmd_context * ctx,
+            mtmd_input_chunks * output,
+            const mtmd_input_part ** parts,
+            size_t n_parts,
+            bool add_special) {
+    for (size_t i = 0; i < n_parts; i++) {
+        if ((parts[i]->text == nullptr) == (parts[i]->bitmap == nullptr)) {
+            LOG_ERR("%s: part %zu must have either text or bitmap set, not both\n", __func__, i);
+            return 1;
+        }
+        if (parts[i]->text != nullptr && parts[i]->text->text == nullptr) {
+            LOG_ERR("%s: part %zu has null text pointer\n", __func__, i);
+            return 1;
+        }
+    }
+    try {
+        mtmd_tokenizer tokenizer(ctx, parts, n_parts, add_special);
+        return tokenizer.tokenize(output);
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: error: %s\n", __func__, e.what());
+        return 2;
+    }
+}
+
 static int32_t mtmd_encode_impl(mtmd_context * ctx, const mtmd_image_tokens * image_tokens, std::vector<float> & out_embd) {
     clip_ctx * ctx_clip = ctx->ctx_v;
     if (!ctx_clip) {
@@ -2132,6 +2176,7 @@ bool mtmd_decode_use_non_causal(const mtmd_context * ctx, const mtmd_input_chunk
         case PROJECTOR_TYPE_GEMMA3:
         case PROJECTOR_TYPE_GEMMA4V:
         case PROJECTOR_TYPE_GEMMA4UV:
+        case PROJECTOR_TYPE_DEEPSEEK4V:
             return true;
         default:
             return false;

tools/mtmd/mtmd.h

@@ -73,6 +73,12 @@ struct mtmd_input_text {
     bool parse_special;
 };
 
+struct mtmd_input_part {
+    // only text or bitmap can be set, not both
+    const struct mtmd_input_text * text;
+    const struct mtmd_bitmap * bitmap;
+};
+
 //
 // C API
 //
@@ -83,6 +89,7 @@ typedef struct mtmd_image_tokens mtmd_image_tokens;
 typedef struct mtmd_input_chunk  mtmd_input_chunk;
 typedef struct mtmd_input_chunks mtmd_input_chunks;
 typedef struct mtmd_input_text   mtmd_input_text;
+typedef struct mtmd_input_part   mtmd_input_part;
 typedef struct mtmd_batch        mtmd_batch;
 
 typedef bool (*mtmd_progress_callback)(float progress, void * user_data);
@@ -276,10 +283,10 @@ struct mtmd_decoder_pos {
 // return relative position (for example, embedding 0 will have position (0, 0, 0); remember to adjust it to the current absolute position)
 MTMD_API struct mtmd_decoder_pos mtmd_image_tokens_get_decoder_pos(const mtmd_image_tokens * image_tokens, llama_pos pos_0, size_t i);
 
-// tokenize an input text prompt and a list of bitmaps (images/audio)
-// the prompt must have the input image marker (default: "<__media__>") in it
+// tokenize an input text prompt and a list of bitmaps (image/audio)
+// the prompt must have the input media marker (default: "<__media__>") in it
 // the default marker is defined by mtmd_default_marker()
-// the marker will be replaced with the image/audio chunk
+// the marker will be replaced with the media chunk
 // for example:
 //   "here is an image: <__media__>\ndescribe it in detail."
 //   this will gives 3 chunks:
@@ -291,13 +298,25 @@ MTMD_API struct mtmd_decoder_pos mtmd_image_tokens_get_decoder_pos(const mtmd_im
 // return values:
 //   0 on success
 //   1 on number of bitmaps not matching the number of markers
-//   2 on image preprocessing error
+//   2 on media preprocessing error
 MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
                                mtmd_input_chunks * output,
                                const mtmd_input_text * text,
                                const mtmd_bitmap ** bitmaps,
                                size_t n_bitmaps);
 
+// same as mtmd_tokenize(), but takes an array of mtmd_input_part
+// use cases:
+// - when you don't want to use media markers (they will be tokenized as normal text)
+// - when you want to control parse_special for each text part
+// note: per-part add_special will be ignored
+// return 1 if a part has both text and bitmap set (or neither)
+MTMD_API int32_t mtmd_tokenize_from_parts(mtmd_context * ctx,
+                                          mtmd_input_chunks * output,
+                                          const mtmd_input_part ** parts,
+                                          size_t n_parts,
+                                          bool add_special);
+
 DEPRECATED(MTMD_API int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens),
            "use mtmd_encode_chunk() instead");
 

tools/server/server-common.cpp

@@ -1062,8 +1062,7 @@ json oaicompat_completion_params_parse(const json & body) {
 static void handle_media(
         std::vector<raw_buffer> & out_files,
         const std::string & url,
-        const std::string & media_path,
-        bool accept_base64_uri) {
+        const std::string & media_path) {
     if (!media_path.empty()) {
         // should already be enforced by arg.cpp, but checking just in case
         GGML_ASSERT(media_path.back() == DIRECTORY_SEPARATOR);
@@ -1104,15 +1103,17 @@ static void handle_media(
         data.assign((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
         out_files.push_back(data);
 
-    } else if (accept_base64_uri && string_starts_with(url, "data:")) {
-        // try to decode base64 image
+    } else if (string_starts_with(url, "data:")) {
+        // try to decode base64 image, video, or audio
         std::vector<std::string> parts = string_split<std::string>(url, /*separator*/ ',');
         if (parts.size() != 2) {
-            throw std::runtime_error("Invalid uri-encoded base64 value");
-        } else if (!string_starts_with(parts[0], "data:image/")) {
-            throw std::runtime_error("Invalid uri format: " + parts[0]);
+            throw std::invalid_argument("Invalid uri-encoded base64 value");
+        } else if (!string_starts_with(parts[0], "data:image/")
+                && !string_starts_with(parts[0], "data:video/")
+                && !string_starts_with(parts[0], "data:audio/")) {
+            throw std::invalid_argument("Invalid uri format: " + parts[0]);
         } else if (!string_ends_with(parts[0], "base64")) {
-            throw std::runtime_error("uri must be base64 encoded");
+            throw std::invalid_argument("uri must be base64 encoded");
         } else {
             auto base64_data = parts[1];
             auto decoded_data = base64_decode(base64_data);
@@ -1219,7 +1220,7 @@ json oaicompat_chat_params_parse(
 
                 json image_url = json_value(p, "image_url", json::object());
                 std::string url = json_value(image_url, "url", std::string());
-                handle_media(out_files, url, opt.media_path, true);
+                handle_media(out_files, url, opt.media_path);
 
                 p["type"] = "media_marker";
                 p["text"] = get_media_marker();
@@ -1234,7 +1235,7 @@ json oaicompat_chat_params_parse(
                 json input_audio = json_value(p, "input_audio", json::object());
                 std::string url  = json_value(input_audio, "data",
                                         json_value(input_audio, "url", std::string()));
-                handle_media(out_files, url, opt.media_path, false);
+                handle_media(out_files, url, opt.media_path);
 
                 p["type"] = "media_marker";
                 p["text"] = get_media_marker();
@@ -1248,7 +1249,7 @@ json oaicompat_chat_params_parse(
                 json input_video = json_value(p, "input_video", json::object());
                 std::string url  = json_value(input_video, "data",
                                         json_value(input_video, "url", std::string()));
-                handle_media(out_files, url, opt.media_path, false);
+                handle_media(out_files, url, opt.media_path);
 
                 p["type"] = "media_marker";
                 p["text"] = get_media_marker();

tools/server/server-context.cpp

@@ -1493,11 +1493,22 @@ private:
                 auto caps = common_chat_templates_get_caps(chat_params.tmpls.get());
                 auto it = params_base.default_template_kwargs.find("preserve_reasoning");
                 bool supported = caps.at("supports_preserve_reasoning");
-                bool enabled = it != params_base.default_template_kwargs.end();
+                bool specified = params_base.preserve_reasoning_specified;
+                // note: the kwarg is enabled by default if not specified explicitly, so check the value
+                bool enabled = it != params_base.default_template_kwargs.end() && it->second == "true";
+                if (supported) {
+                    SRV_TRC("preserve_reasoning kwarg: %s\n",
+                            it == params_base.default_template_kwargs.end() ? "unset (template default)" : it->second.c_str());
+                } else {
+                    SRV_TRC("%s", "preserve_reasoning kwarg: not supported by template\n");
+                }
+                if (supported && !specified) {
+                    SRV_WRN("%s", "chat template supports preserving reasoning, it is enabled by default (may use more tokens, disable via --no-reasoning-preserve)\n");
+                }
                 if (supported && !enabled) {
                     SRV_INF("%s", "chat template supports preserving reasoning, consider enabling it via --reasoning-preserve\n");
                 }
-                if (!supported && enabled) {
+                if (!supported && specified && enabled) {
                     SRV_WRN("%s", "chat template does NOT support preserving reasoning, --reasoning-preserve has no effect\n");
                 }
             }

tools/server/tests/unit/test_vision_api.py

@@ -71,6 +71,7 @@ def test_v1_models_supports_multimodal_capability():
         ("What is this:\n", "malformed",              False, None),
         ("What is this:\n", "https://google.com/404", False, None), # non-existent image
         ("What is this:\n", "https://ggml.ai",        False, None), # non-image data
+        ("What is this:\n", "data:text/html;base64,aGVsbG8=", False, None), # unsupported data uri mime
         # TODO @ngxson : test with multiple images, no images and with audio
     ]
 )