ui : fix ui clipping in mobile due to incorrect height setup (#24605 )

jinja : add count/d/e filter aliases (#24606 )
cli : fix not copying preserved tokens (#24258 )
2026-06-14 17:56:43 +02:00 · 2026-06-14 16:15:51 +02:00 · 2026-06-14 15:07:31 +02:00 · 2026-06-14 11:52:15 +02:00 · 2026-06-14 09:04:46 +02:00 · 2026-06-14 08:27:52 +02:00
26 changed files with 754 additions and 35 deletions
@@ -12,7 +12,7 @@ SYCL:
            - ggml/src/ggml-sycl/**
            - docs/backend/SYCL.md
            - examples/sycl/**
-Nvidia GPU:
+CUDA:
    - changed-files:
        - any-glob-to-any-file:
            - ggml/include/ggml-cuda.h
@@ -783,6 +783,8 @@ jobs:
          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip

  windows-sycl:
+    needs: [check-release]
+    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: windows-2022

@@ -891,6 +893,8 @@ jobs:
          name: llama-bin-win-sycl-x64.zip

  ubuntu-24-sycl:
+    needs: [check-release]
+    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    strategy:
      matrix:
@@ -316,12 +316,22 @@ value filter_expression::execute_impl(context & ctx) {

    JJ_DEBUG("Applying filter to %s", input->type().c_str());

+    auto set_filter_alias = [](auto & filter_id) {
+        if (filter_id == "count") {
+            filter_id = "length";
+        } else if (filter_id == "d") {
+            filter_id = "default";
+        } else if (filter_id == "e") {
+            filter_id = "escape";
+        } else if (filter_id == "trim") {
+            filter_id = "strip";
+        }
+    };
+
    if (is_stmt<identifier>(filter)) {
        auto filter_id = cast_stmt<identifier>(filter)->val;

-        if (filter_id == "trim") {
-            filter_id = "strip"; // alias
-        }
+        set_filter_alias(filter_id);
        JJ_DEBUG("Applying filter '%s' to %s", filter_id.c_str(), input->type().c_str());
        // TODO: Refactor filters so this coercion can be done automatically
        if (!input->is_undefined() && !is_val<value_string>(input) && (
@@ -345,9 +355,7 @@ value filter_expression::execute_impl(context & ctx) {
        }
        auto filter_id = cast_stmt<identifier>(call->callee)->val;

-        if (filter_id == "trim") {
-            filter_id = "strip"; // alias
-        }
+        set_filter_alias(filter_id);
        JJ_DEBUG("Applying filter '%s' with arguments to %s", filter_id.c_str(), input->type().c_str());
        func_args args(ctx);
        for (const auto & arg_expr : call->args) {
@@ -761,9 +769,9 @@ value member_expression::execute_impl(context & ctx) {

        if (is_stmt<slice_expression>(this->property)) {
            auto s = cast_stmt<slice_expression>(this->property);
-            value start_val = s->start_expr ? s->start_expr->execute(ctx) : mk_val<value_int>(0);
-            value stop_val  = s->stop_expr  ? s->stop_expr->execute(ctx)  : mk_val<value_int>(arr_size);
            value step_val  = s->step_expr  ? s->step_expr->execute(ctx)  : mk_val<value_int>(1);
+            value start_val = s->start_expr ? s->start_expr->execute(ctx) : (step_val->as_int() < 0 ? mk_val<value_int>(arr_size - 1) : mk_val<value_int>(0));
+            value stop_val  = s->stop_expr  ? s->stop_expr->execute(ctx)  : (step_val->as_int() < 0 ? mk_val<value_int>(-1) : mk_val<value_int>(arr_size));

            // translate to function call: obj.slice(start, stop, step)
            JJ_DEBUG("Member expression is a slice: start %s, stop %s, step %s",
@@ -90,14 +90,14 @@ static T slice(const T & array, int64_t start, int64_t stop, int64_t step = 1) {
            stop_val = std::min(stop_val, len);
        }
    } else {
-        start_val = len - 1;
+        start_val = start;
        if (start_val < 0) {
-            start_val = std::max(len + start_val, (int64_t)-1);
+            start_val = std::max(len + start_val, (int64_t)0);
        } else {
            start_val = std::min(start_val, len - 1);
        }

-        stop_val = -1;
+        stop_val = stop;
        if (stop_val < -1) {
            stop_val = std::max(len + stop_val, (int64_t)-1);
        } else {
@@ -40,6 +40,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "ChatGLMModel": "chatglm",
    "CodeShellForCausalLM": "codeshell",
    "CogVLMForCausalLM": "cogvlm",
+    "Cohere2MoeForCausalLM": "command_r",
    "Cohere2ForCausalLM": "command_r",
    "CohereForCausalLM": "command_r",
    "DbrxForCausalLM": "dbrx",
@@ -1195,7 +1195,7 @@ class TextModel(ModelBase):
            self.gguf_writer.add_embedding_length(n_embd)
            logger.info(f"gguf: embedding length = {n_embd}")

-        if (n_ff := self.find_hparam(["intermediate_size", "n_inner", "hidden_dim"], optional=True)) is not None:
+        if (n_ff := self.find_hparam(["prefix_dense_intermediate_size", "intermediate_size", "n_inner", "hidden_dim"], optional=True)) is not None:
            self.gguf_writer.add_feed_forward_length(n_ff)
            logger.info(f"gguf: feed forward length = {n_ff}")

@@ -1280,7 +1280,7 @@ class TextModel(ModelBase):
            self.gguf_writer.add_expert_group_used_count(n_group_used)
            logger.info(f"gguf: expert groups used count = {n_group_used}")

-        if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation", "moe_router_activation_func"], optional=True)) is not None:
+        if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation", "moe_router_activation_func", "expert_selection_fn"], optional=True)) is not None:
            if score_func == "sigmoid":
                self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
            elif score_func == "softmax":
@@ -1495,6 +1495,9 @@ class TextModel(ModelBase):
        if chkhsh == "d772b220ace2baec124bed8cfafce0ead7d6c38a4b65ef11261cf9d5d62246d1":
            # ref: https://huggingface.co/CohereLabs/tiny-aya-base
            res = "tiny_aya"
+        if chkhsh == "52df12b4c8d4176e7481aab4b6e8454d1fd0a210a04a574f6d4e067d10e23c3e":
+            # ref: https://huggingface.co/CohereLabs/North-Mini-Code-1.0
+            res = "cohere2moe"
        if chkhsh == "e636dc30a262dcc0d8c323492e32ae2b70728f4df7dfe9737d9f920a282b8aea":
            # ref: https://huggingface.co/Qwen/Qwen1.5-7B
            res = "qwen2"
@@ -1,5 +1,6 @@
 from __future__ import annotations

+import re
 from typing import Iterable, TYPE_CHECKING

 import torch
@@ -55,3 +56,122 @@ class Cohere2Model(TextModel):
            return

        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Cohere2MoeForCausalLM")
+class Cohere2MoeModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.COHERE2MOE
+    _n_main_layers: int | None = None
+    _expert_tensor_re = re.compile(
+        r"model\.layers\.(\d+)\.mlp\.experts\.(\d+)\.(down_proj|gate_proj|up_proj)\.weight"
+    )
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if (n_nextn := int(self.hparams.get("num_nextn_predict_layers", 0) or 0)) > 0 and not self.no_mtp:
+            self.block_count += n_nextn
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+        self._experts: list[dict[str, Tensor]] = [{} for _ in range(self.block_count)]
+
+    def _set_vocab_gpt2(self) -> None:
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        expert_intermediate_size = hparams["intermediate_size"]
+        mlp_layer_types = hparams.get("mlp_layer_types")
+        n_dense_lead = hparams.get("first_k_dense_replace", 0)
+        if mlp_layer_types is not None:
+            n_dense_lead = next((i for i, t in enumerate(mlp_layer_types) if t != "dense"), len(mlp_layer_types))
+
+        super().set_gguf_parameters()
+
+        self.gguf_writer.add_logit_scale(hparams["logit_scale"])
+        self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+        self.gguf_writer.add_sliding_window_pattern([t == "sliding_attention" for t in hparams["layer_types"]])
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_expert_feed_forward_length(expert_intermediate_size)
+        self.gguf_writer.add_leading_dense_block_count(n_dense_lead)
+        self.gguf_writer.add_expert_weights_norm(hparams.get("norm_topk_prob", False))
+        if (num_shared_experts := hparams.get("num_shared_experts", 0)) > 0:
+            if hparams.get("shared_expert_combination_strategy", "average") != "average":
+                raise ValueError("Cohere2 MoE only supports average shared expert combination")
+            self.gguf_writer.add_expert_shared_count(num_shared_experts)
+            self.gguf_writer.add_expert_shared_feed_forward_length(expert_intermediate_size * num_shared_experts)
+        if (n_nextn := hparams.get("num_nextn_predict_layers", 0)) > 0 and not self.no_mtp:
+            self.gguf_writer.add_nextn_predict_layers(n_nextn)
+        self.gguf_writer.add_rope_dimension_count(hparams["head_dim"])
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
+        self._n_main_layers = hparams.get("num_hidden_layers")
+        type(self)._n_main_layers = self._n_main_layers
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item):
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        if cls._n_main_layers is not None:
+            is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
+            if is_mtp and cls.no_mtp:
+                return None
+            if cls.mtp_only and not is_mtp and name not in (
+                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+            ):
+                return None
+
+        return name, gen
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.endswith(".bias"):
+            if torch.any(data_torch != 0):
+                raise ValueError(f"Bias tensor {name!r} is not zero.")
+            logger.debug(f"Skipping bias tensor {name!r}.")
+            return
+
+        if (m := self._expert_tensor_re.fullmatch(name)) is not None:
+            n_experts = self.hparams["num_experts"]
+            layer_idx = int(m.group(1))
+            assert bid is None or bid == layer_idx
+
+            self._experts[layer_idx][name] = data_torch
+
+            expected = {
+                f"model.layers.{layer_idx}.mlp.experts.{xid}.{w_name}.weight"
+                for xid in range(n_experts)
+                for w_name in ("down_proj", "gate_proj", "up_proj")
+            }
+            if expected.issubset(self._experts[layer_idx]):
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{layer_idx}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[layer_idx][ename])
+                        del self._experts[layer_idx][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+                    merged_name = f"model.layers.{layer_idx}.mlp.experts.{w_name}.weight"
+
+                    yield from super().modify_tensors(data_torch, merged_name, layer_idx)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        experts = [k for d in self._experts for k in d.keys()]
+        if len(experts) > 0:
+            raise ValueError(f"Unprocessed experts: {experts}")
@@ -100,6 +100,7 @@ models = [
    {"name": "refact",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
    {"name": "command-r",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
    {"name": "tiny_aya",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereLabs/tiny-aya-base", },
+    {"name": "cohere2moe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereLabs/North-Mini-Code-1.0", },
    {"name": "qwen2",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
    {"name": "olmo",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
    {"name": "dbrx",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
@@ -457,6 +457,7 @@ class MODEL_ARCH(IntEnum):
    XVERSE           = auto()
    COMMAND_R        = auto()
    COHERE2          = auto()
+    COHERE2MOE       = auto()
    DBRX             = auto()
    OLMO             = auto()
    OLMO2            = auto()
@@ -1012,6 +1013,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.XVERSE:           "xverse",
    MODEL_ARCH.COMMAND_R:        "command-r",
    MODEL_ARCH.COHERE2:          "cohere2",
+    MODEL_ARCH.COHERE2MOE:       "cohere2moe",
    MODEL_ARCH.DBRX:             "dbrx",
    MODEL_ARCH.OLMO:             "olmo",
    MODEL_ARCH.OLMO2:            "olmo2",
@@ -2872,6 +2874,33 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_DOWN,
        MODEL_TENSOR.FFN_UP,
    ],
+    MODEL_ARCH.COHERE2MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_GATE_UP_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
+    ],
    MODEL_ARCH.DBRX: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
@@ -16,6 +16,7 @@ set(HF_VERSION        "" CACHE STRING "Version to download (empty = resolve from
 set(HF_ENABLED        "" CACHE STRING "Whether to allow HF Bucket download (ON/OFF)")
 set(BUILD_UI          "" CACHE STRING "Build UI via npm (ON/OFF)")
 set(LLAMA_UI_EMBED    "" CACHE STRING "Path to llama-ui-embed helper")
+set(LLAMA_UI_GZIP     "" CACHE STRING "Apply gzip compress to assets to save bandwidth")

 set(DIST_DIR     "${UI_BINARY_DIR}/dist")
 set(SRC_DIST_DIR "${UI_SOURCE_DIR}/dist")
@@ -225,6 +226,33 @@ function(hf_download version out_var out_resolved)
 endfunction()

 function(emit_files dist_dir)
+    # If gzip is requested, compress every asset into a parallel _gzip/ tree
+    # the structure stays the same; for ex: /abc/def --> /_gzip/abc/def
+    # embed.cpp will check for _gzip and will pick it up
+    if(LLAMA_UI_GZIP AND EXISTS "${dist_dir}/index.html")
+        find_program(GZIP_EXECUTABLE gzip)
+        if(NOT GZIP_EXECUTABLE)
+            message(WARNING "UI: LLAMA_UI_GZIP requested but gzip not found, embedding uncompressed")
+        else()
+            set(gzip_dir "${dist_dir}/_gzip")
+            file(REMOVE_RECURSE "${gzip_dir}")
+            file(GLOB_RECURSE all_files RELATIVE "${dist_dir}" "${dist_dir}/*")
+            foreach(f ${all_files})
+                get_filename_component(dst_dir "${gzip_dir}/${f}" DIRECTORY)
+                file(MAKE_DIRECTORY "${dst_dir}")
+                execute_process(
+                    COMMAND "${GZIP_EXECUTABLE}" -c "${dist_dir}/${f}"
+                    OUTPUT_FILE "${gzip_dir}/${f}"
+                    RESULT_VARIABLE gz_rc
+                )
+                if(NOT gz_rc EQUAL 0)
+                    message(FATAL_ERROR "UI: gzip failed for ${f}")
+                endif()
+            endforeach()
+            message(STATUS "UI: gzip compression applied (${gzip_dir})")
+        endif()
+    endif()
+
    set(args "${UI_CPP}" "${UI_H}")
    if(EXISTS "${dist_dir}/index.html")
        list(APPEND args "${dist_dir}")
@@ -66,6 +66,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_XVERSE,           "xverse"           },
    { LLM_ARCH_COMMAND_R,        "command-r"        },
    { LLM_ARCH_COHERE2,          "cohere2"          },
+    { LLM_ARCH_COHERE2MOE,       "cohere2moe"       },
    { LLM_ARCH_DBRX,             "dbrx"             },
    { LLM_ARCH_OLMO,             "olmo"             },
    { LLM_ARCH_OLMO2,            "olmo2"            },
@@ -71,6 +71,7 @@ enum llm_arch {
    LLM_ARCH_XVERSE,
    LLM_ARCH_COMMAND_R,
    LLM_ARCH_COHERE2,
+    LLM_ARCH_COHERE2MOE,
    LLM_ARCH_DBRX,
    LLM_ARCH_OLMO,
    LLM_ARCH_OLMO2,
@@ -18,6 +18,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
        case LLM_ARCH_GEMMA3:
        case LLM_ARCH_GEMMA3N:
        case LLM_ARCH_COHERE2:
+        case LLM_ARCH_COHERE2MOE:
        case LLM_ARCH_OLMO2:
        case LLM_ARCH_BITNET:
        case LLM_ARCH_T5:
@@ -157,6 +157,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
            return new llama_model_command_r(params);
        case LLM_ARCH_COHERE2:
            return new llama_model_cohere2(params);
+        case LLM_ARCH_COHERE2MOE:
+            return new llama_model_cohere2moe(params);
        case LLM_ARCH_DBRX:
            return new llama_model_dbrx(params);
        case LLM_ARCH_OLMO:
@@ -1467,9 +1469,12 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
    }
    ml.done_getting_tensors();

+    // Tied NVFP4 output is valid when no separate LM-head scale tensors are present.
+    // If sidecar scales exist, the output weight must be an actual output tensor.
    GGML_ASSERT(!(output && tok_embd &&
            strcmp(output->name, tok_embd->name) == 0 &&
-            output->type == GGML_TYPE_NVFP4));
+            output->type == GGML_TYPE_NVFP4 &&
+            (output_s || output_in_s)));
    // populate tensors_by_name
    for (auto & [_, ctx_ptr] : ml.ctx_map) {
        for (auto * cur = ggml_get_first_tensor(ctx_ptr.get()); cur != NULL; cur = ggml_get_next_tensor(ctx_ptr.get(), cur)) {
@@ -1844,6 +1849,7 @@ void llama_model::print_info() const {
        }

        if (arch == LLM_ARCH_MELLUM ||
+                arch == LLM_ARCH_COHERE2MOE ||
                arch == LLM_ARCH_QWEN3MOE ||
                arch == LLM_ARCH_OPENAI_MOE ||
                arch == LLM_ARCH_QWEN3VLMOE ||
@@ -2389,6 +2395,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
        case LLM_ARCH_XVERSE:
        case LLM_ARCH_COMMAND_R:
        case LLM_ARCH_COHERE2:
+        case LLM_ARCH_COHERE2MOE:
        case LLM_ARCH_OLMO:
        case LLM_ARCH_ARCTIC:
        case LLM_ARCH_DEEPSEEK:
@@ -2280,7 +2280,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                clean_spaces = false;
                ignore_merges = true;
            } else if (
-                tokenizer_pre == "tiny_aya") {
+                tokenizer_pre == "tiny_aya" ||
+                tokenizer_pre == "cohere2moe") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_TINY_AYA;
                clean_spaces = false;
            } else if (
@@ -122,9 +122,9 @@ llama_model_cohere2::graph::graph(const llama_model & model, const llm_graph_par
        // feed-forward network
        {
            cur = build_ffn(ffn_inp,
-                    model.layers[il].ffn_up, NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
+                    model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
+                    model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
+                    model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
            cb(cur, "ffn_out", il);
        }
@@ -0,0 +1,443 @@
+#include "models.h"
+
+void llama_model_cohere2moe::load_arch_hparams(llama_model_loader & ml) {
+    const bool found_norm     = ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,     hparams.f_norm_eps,     false);
+    const bool found_norm_rms = ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps, false);
+    if (!found_norm && !found_norm_rms) {
+        throw std::runtime_error("missing Cohere2 MoE norm epsilon");
+    }
+    if (!found_norm_rms) {
+        hparams.f_norm_rms_eps = 0.0f;
+    }
+
+    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(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);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,        hparams.n_layer_nextn, false);
+    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer");
+
+    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+    }
+
+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+    uint32_t swa_period = 4;
+    if (ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false)) {
+        hparams.set_swa_pattern(swa_period, true);
+    } else {
+        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer());
+    }
+
+    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+
+    switch (hparams.n_layer()) {
+        case 49: type = LLM_TYPE_30B_A3B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_cohere2moe::load_arch_tensors(llama_model_loader & ml) {
+    LLAMA_LOAD_LOCALS;
+
+    const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);
+    // Trunk-only: the GGUF declares MTP layers in metadata but the actual MTP
+    // tensors live in a separate file. Mark MTP tensors NOT_REQUIRED so the
+    // trunk loads cleanly.
+    const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
+    const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
+    const int trunk_flags = mtp_only  ? TENSOR_NOT_REQUIRED : 0;
+    const int mtp_flags   = trunk_only ? TENSOR_NOT_REQUIRED : 0;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
+    }
+
+    if (n_expert == 0) {
+        throw std::runtime_error("n_expert must be > 0 for Cohere2Moe");
+    }
+    if (n_expert_used == 0) {
+        throw std::runtime_error("n_expert_used must be > 0 for Cohere2Moe");
+    }
+
+    auto load_block_trunk = [&](int i, int flags) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
+
+        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);
+
+        if (static_cast<uint32_t>(i) < hparams.n_layer_dense_lead) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, flags);
+            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;
+
+            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);
+            create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, flags);
+
+            if (hparams.n_expert_shared > 0) {
+                const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp * hparams.n_expert_shared;
+                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
+                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, flags);
+                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), { n_embd, n_ff_shexp }, flags);
+            }
+        }
+    };
+
+    auto load_block_mtp = [&](int i, int flags) {
+        auto & layer = layers[i];
+
+        // MTP block looks like a full-attention Cohere2 MoE decoder block.
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
+
+        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;
+
+        // Routed experts
+        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);
+        create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, flags);
+
+        if (hparams.n_expert_shared > 0) {
+            const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp * hparams.n_expert_shared;
+
+            // Shared experts
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, flags);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), { n_embd, n_ff_shexp }, flags);
+        }
+
+        // NextN-specific tensors that define the MTP block.
+        layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), { 2 * n_embd, n_embd }, flags);
+        layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), { n_embd },              flags);
+        layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), { n_embd },              flags);
+        layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", i), { n_embd, n_vocab },     TENSOR_NOT_REQUIRED);
+        layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab },     TENSOR_NOT_REQUIRED);
+        layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd },              TENSOR_NOT_REQUIRED);
+    };
+
+    for (int i = 0; i < n_layer; ++i) {
+        load_block_trunk(i, trunk_flags);
+    }
+    // MTP/NextN layers are loaded as extra decoder blocks.
+    for (int i = n_layer; i < n_layer_all; ++i) {
+        load_block_mtp(i, mtp_flags);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_cohere2moe::build_arch_graph(const llm_graph_params & params) const {
+    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
+        return std::make_unique<graph_mtp>(*this, params);
+    }
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_cohere2moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
+
+    const llm_norm_type cohere2moe_norm_type = hparams.f_norm_rms_eps == 0.0f ? LLM_NORM : LLM_NORM_RMS;
+    const float f_logit_scale = hparams.f_logit_scale;
+    ggml_tensor * cur;
+    ggml_tensor * inpL = build_inp_embd(model.tok_embd);
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv_iswa();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass.
+    for (int il = 0; il < n_layer; ++il) {
+        const bool is_swa = hparams.is_swa(il);
+        // Dense-prefix full-attention layers use RoPE; later layers follow the SWA pattern.
+        const bool force_rope = static_cast<uint32_t>(il) < hparams.n_layer_dense_lead;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, cohere2moe_norm_type, il);
+        cb(cur, "attn_norm", il);
+
+        ggml_tensor * ffn_inp = cur;
+
+        {
+            const auto & layer = model.layers[il];
+
+            auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur,
+                    n_embd_head, n_head, n_head_kv, il);
+
+            if (is_swa || force_rope) {
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    layer.wo, layer.wo_b, layer.wo_s,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
+                    1.0f / sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
+            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+
+        ggml_tensor * attn_out = cur;
+
+        const auto & layer = model.layers[il];
+
+        if (layer.ffn_gate_inp == nullptr) {
+            cur = build_ffn(ffn_inp,
+                    layer.ffn_up,   nullptr, layer.ffn_up_s,
+                    layer.ffn_gate, nullptr, layer.ffn_gate_s,
+                    layer.ffn_down, nullptr, layer.ffn_down_s,
+                    nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = build_moe_ffn(ffn_inp,
+                    layer.ffn_gate_inp,
+                    layer.ffn_up_exps,
+                    layer.ffn_gate_exps,
+                    layer.ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, hparams.expert_weights_norm,
+                    hparams.expert_weights_scale,
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il,
+                    nullptr, layer.ffn_gate_up_exps,
+                    layer.ffn_up_exps_s,
+                    layer.ffn_gate_exps_s,
+                    layer.ffn_down_exps_s);
+            cb(cur, "ffn_moe_out", il);
+
+            if (layer.ffn_up_shexp) {
+                ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
+                        layer.ffn_up_shexp,   nullptr, layer.ffn_up_shexp_s,
+                        layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s,
+                        layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s,
+                        nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, cur, ffn_shexp);
+                cur = ggml_scale(ctx0, cur, 0.5f);
+                cb(cur, "ffn_out", il);
+            }
+        }
+
+        cur = ggml_add(ctx0, cur, inpL);
+        cur = ggml_add(ctx0, cur, attn_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, nullptr, cohere2moe_norm_type, -1);
+
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (!cparams.embeddings_nextn_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    if (f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, f_logit_scale);
+    }
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+llama_model_cohere2moe::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    GGML_ASSERT(hparams.n_layer_nextn > 0 && "COHERE2MOE MTP requires n_layer_nextn > 0");
+    GGML_ASSERT(hparams.n_layer_nextn == 1 && "COHERE2MOE MTP currently only supports a single MTP block");
+
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
+
+    const int il = hparams.n_layer();
+    const auto & layer = model.layers[il];
+    GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");
+    GGML_ASSERT(layer.nextn.enorm   && "MTP block missing nextn.enorm");
+    GGML_ASSERT(layer.nextn.hnorm   && "MTP block missing nextn.hnorm");
+    GGML_ASSERT(layer.ffn_gate_inp  && "MTP block missing ffn_gate_inp");
+
+    const llm_norm_type cohere2moe_norm_type = hparams.f_norm_rms_eps == 0.0f ? LLM_NORM : LLM_NORM_RMS;
+
+    // TODO: extract in a common llm_graph_context::build_inp_embd_h()
+    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);
+
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    ggml_set_input(inp->tokens);
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
+    ggml_set_input(inp->embd);
+
+    // TODO: make static using `ggml_build_forward_select()`
+    //       see llm_graph_context::build_inp_embd() for reference
+    ggml_tensor * tok_embd;
+    if (ubatch.token) {
+        ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
+        tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
+    } else {
+        tok_embd = inp->embd;
+    }
+    cb(tok_embd, "mtp_tok_embd", il);
+
+    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
+    ggml_set_input(inp->h);
+    ggml_set_name(inp->h, "mtp_h_input");
+
+    ggml_tensor * h_embd = inp->h;
+
+    res->add_input(std::move(inp));
+
+    ggml_tensor * inp_pos     = build_inp_pos();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, cohere2moe_norm_type, il);
+    cb(h_norm, "mtp_hnorm", il);
+
+    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, cohere2moe_norm_type, il);
+    cb(e_norm, "mtp_enorm", il);
+
+    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0);
+    cb(concat, "mtp_concat", il);
+
+    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);
+    cb(cur, "mtp_eh_proj", il);
+
+    ggml_tensor * inpL = cur;
+
+    cur = build_norm(cur, layer.attn_norm, nullptr, cohere2moe_norm_type, il);
+    cb(cur, "mtp_attn_norm", il);
+    ggml_tensor * ffn_inp = cur;
+
+    auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur, n_embd_head, n_head, n_head_kv, il);
+    ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+    Qcur = ggml_rope_ext(
+            ctx0, Qcur, inp_pos, rope_factors,
+            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+    Kcur = ggml_rope_ext(
+            ctx0, Kcur, inp_pos, rope_factors,
+            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+
+    cb(Qcur, "mtp_Qcur", il);
+    cb(Kcur, "mtp_Kcur", il);
+    cb(Vcur, "mtp_Vcur", il);
+
+    cur = build_attn(inp_attn,
+            layer.wo, layer.wo_b, layer.wo_s,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
+            1.0f / sqrtf(float(n_embd_head)), il);
+    cb(cur, "mtp_attn_out", il);
+
+    ggml_tensor * attn_out = cur;
+
+    cur = build_moe_ffn(ffn_inp,
+            layer.ffn_gate_inp,
+            layer.ffn_up_exps,
+            layer.ffn_gate_exps,
+            layer.ffn_down_exps,
+            nullptr,
+            n_expert, n_expert_used,
+            LLM_FFN_SILU, hparams.expert_weights_norm,
+            hparams.expert_weights_scale,
+            (llama_expert_gating_func_type) hparams.expert_gating_func,
+            il,
+            nullptr, layer.ffn_gate_up_exps,
+            layer.ffn_up_exps_s,
+            layer.ffn_gate_exps_s,
+            layer.ffn_down_exps_s);
+    cb(cur, "mtp_ffn_moe_out", il);
+
+    if (layer.ffn_up_shexp) {
+        ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
+                layer.ffn_up_shexp,   nullptr, layer.ffn_up_shexp_s,
+                layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s,
+                layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s,
+                nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(ffn_shexp, "mtp_ffn_shexp", il);
+
+        cur = ggml_add(ctx0, cur, ffn_shexp);
+        cur = ggml_scale(ctx0, cur, 0.5f);
+        cb(cur, "mtp_ffn_out", il);
+    }
+
+    cur = ggml_add(ctx0, cur, inpL);
+    cur = ggml_add(ctx0, cur, attn_out);
+    cb(cur, "mtp_post_ffn", il);
+
+    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
+            ? layer.nextn.shared_head_norm
+            : model.output_norm;
+    GGML_ASSERT(head_norm_w && "COHERE2MOE MTP: missing both nextn.shared_head_norm and output_norm");
+    cur = build_norm(cur, head_norm_w, nullptr, cohere2moe_norm_type, -1);
+
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    cb(cur, "mtp_shared_head_norm", -1);
+
+    ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;
+    GGML_ASSERT(head_w && "COHERE2MOE MTP: missing LM head (nextn.shared_head_head or model.output)");
+    cur = build_lora_mm(head_w, cur, layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : nullptr);
+
+    if (hparams.f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
+    }
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
@@ -937,6 +937,23 @@ struct llama_model_cohere2 : public llama_model_base {
 };


+struct llama_model_cohere2moe : public llama_model_base {
+    llama_model_cohere2moe(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    struct graph : public llm_graph_context {
+        graph(const llama_model & model, const llm_graph_params & params);
+    };
+
+    struct graph_mtp : public llm_graph_context {
+        graph_mtp(const llama_model & model, const llm_graph_params & params);
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
 struct llama_model_dbrx : public llama_model_base {
    llama_model_dbrx(const struct llama_model_params & params) : llama_model_base(params) {}
    void load_arch_hparams(llama_model_loader & ml) override;
@@ -435,6 +435,24 @@ static void test_expressions(testing & t) {
        "('c', 'b', 'a')"
    );

+    test_template(t, "string slice negative step",
+        "{{ 'abcdef'[::-2] }}",
+        json::object(),
+        "fdb"
+    );
+
+    test_template(t, "string slice negative start and step",
+        "{{ 'abcdef'[-1:1:-1] }}",
+        json::object(),
+        "fedc"
+    );
+
+    test_template(t, "string slice negative start, stop and step",
+        "{{ 'abcdef'[-1:-5:-1] }}",
+        json::object(),
+        "fedc"
+    );
+
    test_template(t, "arithmetic",
        "{{ (a + b) * c }}",
        {{"a", 2}, {"b", 3}, {"c", 4}},
@@ -583,8 +601,8 @@ static void test_filters(testing & t) {
        "hello jinja"
    );

-    test_template(t, "length list",
-        "{{ items|length }}",
+    test_template(t, "length (count alias) list",
+        "{{ items|count }}",
        {{"items", json::array({1, 2, 3})}},
        "3"
    );
@@ -693,8 +711,8 @@ static void test_filters(testing & t) {
        "fallback"
    );

-    test_template(t, "default with falsy value",
-        "{{ ''|default('fallback', true) }}",
+    test_template(t, "default (d alias) with falsy value",
+        "{{ ''|d('fallback', true) }}",
        json::object(),
        "fallback"
    );
@@ -185,7 +185,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
        ms.add_kv(LLM_KV_ROPE_FREQ_BASE_SWA,              10000.0f);
        // SWA pattern: every 5th layer is full attention (matches E2B layer_types)
        ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, uint32_t(5));
-    } else if (arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35) {
+    } else if (arch == LLM_ARCH_COHERE2MOE || arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_STEP35) {
        std::vector<uint32_t> pattern;
        pattern.reserve(n_layer);
        for (uint32_t il = 0; il < n_layer; il++) {
@@ -322,6 +322,7 @@ static std::vector<float> get_logits(
 static bool moe_mandatory(const llm_arch arch) {
    switch (arch) {
        case LLM_ARCH_LLAMA4:
+        case LLM_ARCH_COHERE2MOE:
        case LLM_ARCH_GROK:
        case LLM_ARCH_QWEN2MOE:
        case LLM_ARCH_QWEN3MOE:
@@ -97,11 +97,18 @@ struct cli_context {
                task.params.chat_parser_params.parser.load(chat_params.parser);
            }

+            // Copy the preserved tokens into the sampling params
+            const llama_vocab * vocab = llama_model_get_vocab(
+                llama_get_model(ctx_server.get_llama_context()));
+            for (const auto & token : chat_params.preserved_tokens) {
+                auto ids = common_tokenize(vocab, token, false, true);
+                if (ids.size() == 1) {
+                    task.params.sampling.preserved_tokens.insert(ids[0]);
+                }
+            }
+
            // reasoning budget sampler
            if (!chat_params.thinking_end_tag.empty()) {
-                const llama_vocab * vocab = llama_model_get_vocab(
-                    llama_get_model(ctx_server.get_llama_context()));
-
                task.params.sampling.reasoning_budget_tokens = defaults.sampling.reasoning_budget_tokens;
                task.params.sampling.generation_prompt = chat_params.generation_prompt;

@@ -319,8 +319,26 @@ bool server_http_context::init(const common_params & params) {
            }
        } else {
 #if defined(LLAMA_UI_HAS_ASSETS)
+            static auto handle_gzip_header = [](const httplib::Request & req, httplib::Response & res) {
+                if (!llama_ui_use_gzip()) {
+                    // no gzip build, skip
+                    return true;
+                }
+                if (req.get_header_value("Accept-Encoding").find("gzip") == std::string::npos) {
+                    res.status = 415; // unsupported media type
+                    res.set_content("Error: gzip is not supported by this browser", "text/plain");
+                    return false;
+                } else {
+                    res.set_header("Content-Encoding", "gzip");
+                }
+                return true;
+            };
+
            auto serve_asset_cached = [](const std::string & name, bool isolation) {
                return [name, isolation](const httplib::Request & req, httplib::Response & res) {
+                    if (!handle_gzip_header(req, res)) {
+                        return true; // returns error message
+                    }
                    const llama_ui_asset * a = llama_ui_find_asset(name);
                    if (!a) { res.status = 404; return false; }
                    res.set_header("ETag", a->etag);
@@ -340,7 +358,10 @@ bool server_http_context::init(const common_params & params) {
            };

            auto serve_asset_nocache = [](const std::string & name) {
-                return [name](const httplib::Request & /*req*/, httplib::Response & res) {
+                return [name](const httplib::Request & req, httplib::Response & res) {
+                    if (!handle_gzip_header(req, res)) {
+                        return true; // returns error message
+                    }
                    const llama_ui_asset * a = llama_ui_find_asset(name);
                    if (!a) {
                        res.status = 404;
@@ -1,6 +1,7 @@
 set(TARGET llama-ui)

 set(LLAMA_UI_HF_BUCKET "ggml-org/llama-ui" CACHE STRING "Hugging Face bucket name for prebuilt UI assets")
+set(LLAMA_UI_GZIP      ON                  CACHE BOOL   "Apply gzip compress to assets to save bandwidth")

 # Backward compat: forward old var to new one
 if(DEFINED LLAMA_BUILD_WEBUI)
@@ -83,6 +84,7 @@ add_custom_target(llama-ui-assets ALL
        "-DHF_ENABLED=${LLAMA_USE_PREBUILT_UI}"
        "-DBUILD_UI=${LLAMA_BUILD_UI}"
        "-DLLAMA_UI_EMBED=${LLAMA_UI_EMBED_EXE}"
+        "-DLLAMA_UI_GZIP=${LLAMA_UI_GZIP}"
        -P "${PROJECT_SOURCE_DIR}/scripts/ui-assets.cmake"
    COMMENT "Provisioning UI assets"
    VERBATIM
@@ -145,12 +145,15 @@ int main(int argc, char ** argv) {
        return 1;
    }

-    const std::string out_cpp = argv[1];
-    const std::string out_h   = argv[2];
-    const std::string in_dir  = argv[3];
+    const std::string out_cpp   = argv[1];
+    const std::string out_h     = argv[2];
+    const std::string asset_dir = (argc >= 4) ? argv[3] : std::string();
+
+    const bool        use_gzip = !asset_dir.empty() && std::filesystem::exists(asset_dir + "/_gzip");
+    const std::string in_dir   = use_gzip ? (asset_dir + "/_gzip") : asset_dir;

    std::vector<asset_entry> assets;
-    if (argc == 4) {
+    if (!in_dir.empty()) {
        const std::filesystem::path dir = in_dir;

        std::error_code ec;
@@ -238,7 +241,8 @@ int main(int argc, char ** argv) {
        "    std::string           etag;\n"
        "    std::string           type;\n"
        "};\n\n"
-        "const llama_ui_asset * llama_ui_find_asset(const std::string & name);\n";
+        "const llama_ui_asset * llama_ui_find_asset(const std::string & name);\n"
+        "bool llama_ui_use_gzip();\n";
    h += fmt("const std::array<llama_ui_asset, %d> & llama_ui_get_assets();\n", n_assets);

    std::string cpp;
@@ -294,6 +298,7 @@ int main(int argc, char ** argv) {
            "    return empty;\n"
            "}\n";
    }
+    cpp += fmt("bool llama_ui_use_gzip() { return %s; }\n", use_gzip ? "true" : "false");

    bool ok = true;
    ok = write_if_different(out_h,   h)   && ok;
@@ -41,7 +41,7 @@
 	data-slot="sidebar-wrapper"
 	style="--sidebar-width: {sidebar.sidebarWidth}; --sidebar-min-width: {SIDEBAR_MIN_WIDTH}; --sidebar-max-width: {SIDEBAR_MAX_WIDTH}; --sidebar-width-icon: {SIDEBAR_WIDTH_ICON}; {style}"
 	class={cn(
-		'group/sidebar-wrapper flex min-h-svh w-full has-data-[variant=inset]:bg-sidebar',
+		'group/sidebar-wrapper flex flex-col min-h-svh w-full has-data-[variant=inset]:bg-sidebar',
 		className
 	)}
 	bind:this={ref}
@@ -312,7 +312,7 @@
 	/>

 	<Sidebar.Provider bind:open={sidebarOpen}>
-		<div class="flex h-screen w-full">
+		<div class="flex h-full w-full grow">
 			<Sidebar.Root variant="floating" class="h-full"
 				><SidebarNavigation bind:this={chatSidebar} /></Sidebar.Root
 			>
Author	SHA1	Message	Date
Amos Wong	1fd6dfe9f3	ui : fix ui clipping in mobile due to incorrect height setup (#24605 )	2026-06-14 16:15:51 +02:00
Sigbjørn Skjæret	acd79d603c	jinja : add count/d/e filter aliases (#24606 )	2026-06-14 15:07:31 +02:00
Michael Wand	6e14286eda	cli : fix not copying preserved tokens (#24258 )	2026-06-14 11:52:15 +02:00
Bartowski	8ed274ef46	Add cohere2moe to llama-vocab for TINY_AYA (#24601 )	2026-06-14 09:04:46 +02:00
Sigbjørn Skjæret	46722116b9	ci : use CUDA label for cuda backend (#24594 )	2026-06-14 08:27:52 +02:00
Sigbjørn Skjæret	c2ba3e47a2	add sycl to check-release (#24583 )	2026-06-14 09:42:26 +08:00
Aldehir Rojas	53bd47ea5b	ui : fix llama-ui-embed crash when no asset dir is given (#24597 )	2026-06-13 17:53:30 -05:00
Michael Wand	4988f6e866	Add arch support for cohere2-MoE (#24260 ) * Add arch support for cohere2-MoE * Removed redundant gating_func checks * Changed ffn lookup to prefer prefix_dense_intermediate_size * Renamed arch to cohere2moe * Removed redundant lmhead check and chat template changes * Removed lm_head.weight check from modify tensors, load output tensor not required, fallback to token_embd.weight * Changed to (routed+shared)0.5 for shared expert combined avg fixed sliding_window_pattern issue and pattern * Fixed transformers crash 'first_k_dense_replace' error * Remove comment * Removed cohere2-moe as a tokenizer type and kept as tiny_aya. Renamed North-Mini-Code-1.0. * Fixed MTP fail, changed to use iSWA * Fixed remaining todos: cohere2moe renamed, changed swa parsing to use get_key_or_arr, removed extra get_arr use * Force metadata usage Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Remove Cohere2 checkpoint comment Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Remove MTP comment Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Regenerate cohere2moe tokenizer hash * Add cohere2moe to Llama Model Saver supported list * Check for zerobios tensors and add support for Command to use LayerNorm * Map expert_selection_fn to sigmoid in base.py instead of command.py * use bools for foundnorm/foundnormrms Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>	2026-06-13 19:49:00 +02:00
Sigbjørn Skjæret	f05cf4676a	jinja : fix negative step slice with start/stop values (#24580 )	2026-06-13 18:28:40 +02:00
Xuan-Son Nguyen	e8067a8b36	ui: build-time gzip compression (#24571 ) * ui: keep original file name and path * fix nocache * ui: build-time gzip compression	2026-06-13 16:57:27 +02:00