llama : add --n-cpu-moe option (#15077)

* llama : add --n-cpu-moe option

Keeps the MoE weights of the first N layers in the CPU

.github/workflows/build.yml

@@ -159,31 +159,15 @@ jobs:
       - name: Dawn Dependency
         id: dawn-depends
         run: |
-          ARTIFACTS_JSON=$(curl -s -L \
-            -H "Accept: application/vnd.github+json" \
-            -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
-            -H "X-GitHub-Api-Version: 2022-11-28" \
-            "https://api.github.com/repos/google/dawn/actions/artifacts")
-          echo "Finding latest macos-latest-Release artifact..."
-          DOWNLOAD_URL=$(echo "$ARTIFACTS_JSON" | jq -r '.artifacts
-            | sort_by(.created_at)
-            | reverse
-            | map(select(.name | test("macos-latest-Release$")))
-            | .[0].archive_download_url')
-          if [ "$DOWNLOAD_URL" = "null" ] || [ -z "$DOWNLOAD_URL" ]; then
-            echo "No suitable Dawn artifact found!"
-            exit 1
-          fi
-          echo "Downloading from: $DOWNLOAD_URL"
-          curl -L \
-            -H "Accept: application/vnd.github+json" \
-            -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
-            -o artifact.zip "$DOWNLOAD_URL"
-          unzip artifact.zip
+          DAWN_VERSION="v1.0.0"
+          DAWN_OWNER="reeselevine"
+          DAWN_REPO="dawn"
+          DAWN_ASSET_NAME="Dawn-a1a6b45cced25a3b7f4fb491e0ae70796cc7f22b-macos-latest-Release.tar.gz"
+          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
+          curl -L -o artifact.tar.gz \
+            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
           mkdir dawn
-          tar_file=$(find . -name '*.tar.gz' | head -n 1)
-          echo "Extracting: $tar_file"
-          tar -xvf "$tar_file" -C dawn --strip-components=1
+          tar -xvf artifact.tar.gz -C dawn --strip-components=1
 
       - name: Build
         id: cmake_build
@@ -433,31 +417,15 @@ jobs:
         id: dawn-depends
         run: |
           sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
-          ARTIFACTS_JSON=$(curl -s -L \
-            -H "Accept: application/vnd.github+json" \
-            -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
-            -H "X-GitHub-Api-Version: 2022-11-28" \
-            "https://api.github.com/repos/google/dawn/actions/artifacts")
-          echo "Finding latest ubuntu-latest-Release artifact..."
-          DOWNLOAD_URL=$(echo "$ARTIFACTS_JSON" | jq -r '.artifacts
-            | sort_by(.created_at)
-            | reverse
-            | map(select(.name | test("ubuntu-latest-Release$")))
-            | .[0].archive_download_url')
-          if [ "$DOWNLOAD_URL" = "null" ] || [ -z "$DOWNLOAD_URL" ]; then
-            echo "No suitable Dawn artifact found!"
-            exit 1
-          fi
-          echo "Downloading from: $DOWNLOAD_URL"
-          curl -L \
-            -H "Accept: application/vnd.github+json" \
-            -H "Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}" \
-            -o artifact.zip "$DOWNLOAD_URL"
-          unzip artifact.zip
+          DAWN_VERSION="v1.0.0"
+          DAWN_OWNER="reeselevine"
+          DAWN_REPO="dawn"
+          DAWN_ASSET_NAME="Dawn-a1a6b45cced25a3b7f4fb491e0ae70796cc7f22b-ubuntu-latest-Release.tar.gz"
+          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
+          curl -L -o artifact.tar.gz \
+            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
           mkdir dawn
-          tar_file=$(find . -name '*.tar.gz' | head -n 1)
-          echo "Extracting: $tar_file"
-          tar -xvf "$tar_file" -C dawn --strip-components=1
+          tar -xvf artifact.tar.gz -C dawn --strip-components=1
 
       - name: Build
         id: cmake_build

common/arg.cpp

@@ -24,6 +24,7 @@
 #include <cstdarg>
 #include <filesystem>
 #include <fstream>
+#include <list>
 #include <regex>
 #include <set>
 #include <string>
@@ -2375,20 +2376,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                     }
                     throw std::invalid_argument("unknown buffer type");
                 }
-                // FIXME: this leaks memory
-                params.tensor_buft_overrides.push_back({strdup(tensor_name.c_str()), buft_list.at(buffer_type)});
+                // keep strings alive and avoid leaking memory by storing them in a static vector
+                static std::list<std::string> buft_overrides;
+                buft_overrides.push_back(tensor_name);
+                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), buft_list.at(buffer_type)});
             }
         }
     ));
     add_opt(common_arg(
-        {"--cpu-moe"},
-        "use CPU for Mixture of Experts (MoE) weights",
+        {"--cpu-moe", "-cmoe"},
+        "keep all Mixture of Experts (MoE) weights in the CPU",
         [](common_params & params) {
-            params.tensor_buft_overrides.push_back({"\\.ffn_up_exps\\.weight$",   ggml_backend_cpu_buffer_type()});
-            params.tensor_buft_overrides.push_back({"\\.ffn_down_exps\\.weight$", ggml_backend_cpu_buffer_type()});
-            params.tensor_buft_overrides.push_back({"\\.ffn_gate_exps\\.weight$", ggml_backend_cpu_buffer_type()});
+            params.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
         }
     ).set_env("LLAMA_ARG_CPU_MOE"));
+    add_opt(common_arg(
+        {"--n-cpu-moe", "-ncmoe"}, "N",
+        "keep the Mixture of Experts (MoE) weights of the first N layers in the CPU",
+        [](common_params & params, int value) {
+            if (value < 0) {
+                throw std::invalid_argument("invalid value");
+            }
+            for (int i = 0; i < value; ++i) {
+                // keep strings alive and avoid leaking memory by storing them in a static vector
+                static std::list<std::string> buft_overrides;
+                buft_overrides.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
+                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), ggml_backend_cpu_buffer_type()});
+            }
+        }
+    ).set_env("LLAMA_ARG_N_CPU_MOE"));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -2647,6 +2663,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.n_out_freq = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
+    add_opt(common_arg(
+        {"--output-format"}, "{gguf,dat}",
+        string_format("output format for imatrix file (default: %s)", params.imat_dat > 0 ? "dat" : "gguf"),
+        [](common_params & params, const std::string & value) {
+            /**/ if (value == "gguf") { params.imat_dat = -1; }
+            else if (value == "dat")  { params.imat_dat = 1;  }
+            else { throw std::invalid_argument("invalid output format"); }
+        }
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--save-frequency"}, "N",
         string_format("save an imatrix copy every N iterations (default: %d)", params.n_save_freq),

common/common.h

@@ -439,6 +439,7 @@ struct common_params {
     int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
+    int8_t  imat_dat    =  0; // whether the legacy imatrix.dat format should be output (gguf <= 0 < dat)
 
     bool process_output  = false; // collect data for the output tensor
     bool compute_ppl     = true;  // whether to compute perplexity

convert_hf_to_gguf.py

@@ -678,6 +678,9 @@ class TextModel(ModelBase):
         if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
             # ref: https://huggingface.co/THUDM/glm-4-9b-hf
             res = "glm4"
+        if chkhsh == "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902":
+            # ref: https://huggingface.co/zai-org/GLM-4.5-Air
+            res = "glm4"
         if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
             # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
             res = "minerva-7b"
@@ -852,6 +855,9 @@ class TextModel(ModelBase):
         if chkhsh == "2085e1638f6c377a0aa4ead21b27bb4cb941bf800df86ed391011769c1758dfb":
             # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B
             res = "exaone4"
+        if chkhsh == "a1e163ecab2e718a4c829d1148b6e86824ec36163bb71941c3dca9cd5ac25756":
+            # ref: https://huggingface.co/JetBrains/Mellum-4b-base
+            res = "mellum"
 
         if res is None:
             logger.warning("\n")
@@ -6059,6 +6065,7 @@ class DeepseekModel(TextModel):
 
 @ModelBase.register("DeepseekV2ForCausalLM")
 @ModelBase.register("DeepseekV3ForCausalLM")
+@ModelBase.register("KimiVLForConditionalGeneration")
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
@@ -6161,6 +6168,13 @@ class DeepseekV2Model(TextModel):
     _experts: list[dict[str, Tensor]] | None = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # skip vision tensors and remove "language_model." for Kimi-VL
+        if "vision_tower" in name or "multi_modal_projector" in name:
+            return []
+
+        if name.startswith("language_model."):
+            name = name.replace("language_model.", "")
+
         # rename e_score_correction_bias tensors
         if name.endswith("e_score_correction_bias"):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")
@@ -6685,6 +6699,139 @@ class Glm4Model(TextModel):
         return super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Glm4MoeForCausalLM")
+class Glm4MoeModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.GLM4_MOE
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # GLM4_MOE has num_hidden_layers + 1 actual layers (including NextN layer)
+        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_vocab(self):
+        from transformers import AutoTokenizer
+
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        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 tokens
+        # Note: Using <|endoftext|> (151329) for eot causes endless generation
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # 151331
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # 151336
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) # 151329
+        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
+
+        # Patch broken chat template
+        if isinstance(special_vocab.chat_template, str) and "visible_text(m.content).endswith" in special_vocab.chat_template:
+            special_vocab.chat_template = special_vocab.chat_template.replace(
+                """{{ visible_text(m.content) }}\n{{- '/nothink' if (enable_thinking is defined and not enable_thinking and not visible_text(m.content).endswith("/nothink")) else '' -}}""",
+                """{% set content = visible_text(m.content) %}{{ content }}\n{{- '/nothink' if (enable_thinking is defined and not enable_thinking and not content.endswith("/nothink")) else '' -}}""")
+
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if (rope_dim := self.hparams.get("head_dim")) is None:
+            rope_dim = (
+                self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+            )
+        self.gguf_writer.add_rope_dimension_count(
+            int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5))
+        )
+
+        # MoE parameters - Use only routed expert count (shared experts handled separately)
+        if (n_routed_experts := self.hparams.get("n_routed_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_routed_experts)
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+        if (n_shared_experts := self.hparams.get("n_shared_experts")) is not None:
+            self.gguf_writer.add_expert_shared_count(n_shared_experts)
+        if (first_k_dense_replace := self.hparams.get("first_k_dense_replace")) is not None:
+            self.gguf_writer.add_leading_dense_block_count(first_k_dense_replace)
+
+        # Expert gating function (sigmoid for GLM4_MOE)
+        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+
+        # Routed scaling factor
+        if (routed_scaling_factor := self.hparams.get("routed_scaling_factor")) is not None:
+            self.gguf_writer.add_expert_weights_scale(routed_scaling_factor)
+
+        # Normalise topk probabilities
+        if (norm_topk_prob := self.hparams.get("norm_topk_prob")) is not None:
+            self.gguf_writer.add_expert_weights_norm(norm_topk_prob)
+
+        # NextN/MTP prediction layers
+        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+            self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(
+        self, data_torch: Tensor, name: str, bid: int | None
+    ) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("model.visual."):  # ignore visual part
+            return []
+        elif name.startswith("model.language_model."):
+            name = name.replace("language_model.", "")  # for multimodal variants
+
+        # Handle main token embedding (but not layer-specific NextN embeddings)
+        if name == "model.embed_tokens.weight" and ".layers." not in name:
+            return [(self.map_tensor_name("token_embd.weight"), data_torch)]
+
+        # Handle routed experts
+        if name.find("mlp.experts") != -1:
+            n_experts = self.hparams["n_routed_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        new_name = self.map_tensor_name(name)
+
+        return [(new_name, data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CHATGLM

convert_hf_to_gguf_update.py

@@ -138,6 +138,7 @@ models = [
     {"name": "midm-2.0",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
     {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
+    {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -146,6 +147,7 @@ pre_computed_hashes = [
     {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b"},
     {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"},
     {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"},
+    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.5-Air", "chkhsh": "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902"},
     {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
     {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
     {"name": "hunyuan-dense", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-4B-Instruct", "chkhsh": "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6"},

ggml/CMakeLists.txt

@@ -39,8 +39,9 @@ if (WIN32)
     set(CMAKE_SHARED_MODULE_PREFIX  "")
 endif()
 
-option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
-option(GGML_BACKEND_DL   "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
+option(BUILD_SHARED_LIBS           "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+option(GGML_BACKEND_DL             "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
+set(GGML_BACKEND_DIR "" CACHE PATH "ggml: directory to load dynamic backends from (requires GGML_BACKEND_DL")
 
 #
 # option list

ggml/cmake/ggml-config.cmake.in

@@ -106,7 +106,7 @@ if(NOT TARGET ggml::ggml)
 
     find_library(GGML_LIBRARY ggml
         REQUIRED
-        HINTS ${GGML_LIB_DIR}
+        HINTS ${GGML_LIB_DIR} ${GGML_BACKEND_DIR}
         NO_CMAKE_FIND_ROOT_PATH)
 
     add_library(ggml::ggml UNKNOWN IMPORTED)

ggml/src/CMakeLists.txt

@@ -214,6 +214,13 @@ add_library(ggml
             ggml-backend-reg.cpp)
 add_library(ggml::ggml ALIAS ggml)
 
+if (GGML_BACKEND_DIR)
+    if (NOT GGML_BACKEND_DL)
+        message(FATAL_ERROR "GGML_BACKEND_DIR requires GGML_BACKEND_DL")
+    endif()
+    target_compile_definitions(ggml PUBLIC GGML_BACKEND_DIR="${GGML_BACKEND_DIR}")
+endif()
+
 target_link_libraries(ggml PUBLIC ggml-base)
 
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
@@ -227,7 +234,11 @@ function(ggml_add_backend_library backend)
         set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
         target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
         add_dependencies(ggml ${backend})
-        install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
+        if (GGML_BACKEND_DIR)
+            install(TARGETS ${backend} LIBRARY DESTINATION ${GGML_BACKEND_DIR})
+        else()
+            install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
+        endif()
     else()
         add_library(${backend} ${ARGN})
         target_link_libraries(ggml PUBLIC ${backend})

ggml/src/ggml-backend-reg.cpp

@@ -498,6 +498,9 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
 
     std::vector<fs::path> search_paths;
     if (user_search_path == nullptr) {
+#ifdef GGML_BACKEND_DIR
+        search_paths.push_back(fs::u8path(GGML_BACKEND_DIR));
+#endif
         // default search paths: executable directory, current directory
         search_paths.push_back(get_executable_path());
         search_paths.push_back(fs::current_path());

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

@@ -2046,8 +2046,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
 
     backend_ctx->adreno_cl_compiler_version = get_adreno_cl_compiler_version(driver_version);
     backend_ctx->has_vector_subgroup_broadcast =
-        backend_ctx->adreno_cl_compiler_version.major >= 47 ||
-        backend_ctx->adreno_cl_compiler_version.major == 17;
+        (backend_ctx->adreno_cl_compiler_version.type == E031 && backend_ctx->adreno_cl_compiler_version.major >= 47) ||
+        (backend_ctx->adreno_cl_compiler_version.type == DX   && backend_ctx->adreno_cl_compiler_version.major >= 17);
     GGML_LOG_INFO("ggml_opencl: vector subgroup broadcast support: %s\n",
         backend_ctx->has_vector_subgroup_broadcast ? "true" : "false");
 

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

@@ -3096,6 +3096,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
         uint32_t conv2d_SHMEM_PAD = 4;
         bool conv2d_UNROLL = true;
 
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+        if (device->coopmat2) {
+            conv2d_SHMEM_PAD = 8; // 8 float16_t
+        }
+#endif
+
         if (device->vendor_id == VK_VENDOR_ID_INTEL) {
             conv2d_SHMEM_PAD = 0;
             conv2d_UNROLL = false;
@@ -3154,6 +3160,16 @@ static void ggml_vk_load_shaders(vk_device& device) {
         std::array<uint32_t, 3> wg_denoms = { conv2d_BS_K, conv2d_BS_NPQ, 1 };
         std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };
 
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+        if (device->coopmat2) {
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_cm2_len, conv2d_f32_cm2_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_cm2_len, conv2d_f16_f32_cm2_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+        } else
+#endif
         if (conv2d_UNROLL) {
             ggml_vk_create_pipeline(
                 device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_unroll_len, conv2d_f32_unroll_data, "main", 3,

ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp

@@ -1,6 +1,11 @@
 #version 450
 
 #extension GL_EXT_control_flow_attributes : enable
+#ifdef COOPMAT2
+#extension GL_NV_cooperative_matrix2 : enable
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_KHR_memory_scope_semantics : enable
+#endif
 
 #ifdef USE_COLLECTIVES
 #    extension GL_KHR_shader_subgroup_shuffle : enable
@@ -91,6 +96,12 @@ uint32_t n_elems_out = K * NPQ;
 // Number of blocktiles per input
 uint32_t NB_CRS = splitWork(CRS, BS_CRS);
 
+#ifdef COOPMAT2
+#define SHMEM_TYPE float16_t
+#else
+#define SHMEM_TYPE float
+#endif
+
 const uint32_t Ash_stride = BS_CRS + SHMEM_PAD;
 const uint32_t Bsh_stride = BS_NPQ + SHMEM_PAD;
 
@@ -100,8 +111,8 @@ const uint32_t Bsh_numel = BS_CRS * BS_NPQ;
 const uint32_t Ash_len = BS_K * Ash_stride;
 const uint32_t Bsh_len = BS_CRS * Bsh_stride;
 
-shared float Ash[Ash_len];  // K x CRS
-shared float Bsh[Bsh_len];  // CRS x NPQ
+shared SHMEM_TYPE Ash[Ash_len];  // K x CRS
+shared SHMEM_TYPE Bsh[Bsh_len];  // CRS x NPQ
 
 // Threadtile sizes
 const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
@@ -110,10 +121,6 @@ const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
 const uint32_t NT_K   = BS_K / TS_K;
 const uint32_t NT_NPQ = BS_NPQ / TS_NPQ;
 
-float regA[TS_K];
-float regB[TS_NPQ];
-float regC[TS_K][TS_NPQ];
-
 /*
 Compute
 KxCRS @ CRSxNPQ = K x NPQ
@@ -145,12 +152,36 @@ uint fastdiv(uint n, uint mp, uint L) {
     return (msbs + n) >> L;
 }
 
+#ifdef COOPMAT2
+#define ACC_TYPE float16_t
+
+ACC_TYPE perElemOpStore(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem)
+{
+    uint32_t K_idx   = B_idx_K * BS_K + r;
+    uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + c;
+    uint32_t N_idx   = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
+    uint32_t OH_idx  = fastdiv(NPQ_idx - N_idx * p.OH * p.OW, p.OWmp, p.OWL); // divide by p.OW;
+    uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
+    uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
+    if (K_idx < K && NPQ_idx < NPQ) {
+        dst_data[dst_idx] = D_TYPE(elem);
+    }
+    return elem;
+}
+#endif
+
 void main() {
+#ifdef COOPMAT2
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator> matC;
+    matC = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator>(0.0);
+#else
+    float regC[TS_K][TS_NPQ];
     for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
         for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
             regC[T_ly][T_lx] = 0.0;
         }
     }
+#endif
     /* Advance block in CRS dim */
     for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
         uint32_t CRS_idx_a;
@@ -199,7 +230,7 @@ void main() {
             if (K_idx >= K || CRS_idx_a >= CRS) {
                 val = 0.0;
             }
-            Ash[B_ly * Ash_stride + B_lx] = val;
+            Ash[B_ly * Ash_stride + B_lx] = SHMEM_TYPE(val);
         }
         /* Load input to B_block: (BS_CRS x BS_NPQ) */
         UNROLL for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
@@ -244,11 +275,21 @@ void main() {
             if (CRS_idx_b >= CRS || NPQ_idx >= NPQ || H_idx < 0 || H_idx >= p.H || W_idx < 0 || W_idx >= p.W) {
                 val = 0.0;
             }
-            Bsh[B_ly * Bsh_stride + B_lx] = val;
+            Bsh[B_ly * Bsh_stride + B_lx] = SHMEM_TYPE(val);
         }
         barrier();
+#ifdef COOPMAT2
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_K, BS_CRS, gl_MatrixUseA> matA;
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_CRS, BS_NPQ, gl_MatrixUseB> matB;
+
+        coopMatLoad(matA, Ash, 0, Ash_stride, gl_CooperativeMatrixLayoutRowMajor);
+        coopMatLoad(matB, Bsh, 0, Bsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+        matC = coopMatMulAdd(matA, matB, matC);
+#else
         if (T_y * TS_K < K) {
             UNROLL for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+                float regA[TS_K];
+                float regB[TS_NPQ];
                 for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
                     regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
                 }
@@ -262,9 +303,13 @@ void main() {
                 }
             }
         }
+#endif
         barrier();
     }
     /* Save C* */
+#ifdef COOPMAT2
+    coopMatPerElementNV(matC, matC, perElemOpStore);
+#else
     if (T_y * TS_K < K) {
         for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
             for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
@@ -280,4 +325,5 @@ void main() {
             }
         }
     }
+#endif
 }

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

@@ -661,6 +661,11 @@ void process_shaders() {
     string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
     string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
 
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
+    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
+#endif
+
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
 

gguf-py/gguf/constants.py

@@ -105,6 +105,7 @@ class Keys:
         EXPERT_WEIGHTS_NORM               = "{arch}.expert_weights_norm"
         EXPERT_GATING_FUNC                = "{arch}.expert_gating_func"
         MOE_EVERY_N_LAYERS                = "{arch}.moe_every_n_layers"
+        NEXTN_PREDICT_LAYERS              = "{arch}.nextn_predict_layers"
         POOLING_TYPE                      = "{arch}.pooling_type"
         LOGIT_SCALE                       = "{arch}.logit_scale"
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
@@ -357,6 +358,7 @@ class MODEL_ARCH(IntEnum):
     DEEPSEEK2        = auto()
     CHATGLM          = auto()
     GLM4             = auto()
+    GLM4_MOE         = auto()
     BITNET           = auto()
     T5               = auto()
     T5ENCODER        = auto()
@@ -614,6 +616,13 @@ class MODEL_TENSOR(IntEnum):
     A_MMPROJ_FC          = auto()
     A_MM_NORM_PRE        = auto()
     A_MM_NORM_MID        = auto()
+    # nextn/mtp
+    NEXTN_EH_PROJ        = auto()
+    NEXTN_EMBED_TOKENS   = auto()
+    NEXTN_ENORM          = auto()
+    NEXTN_HNORM          = auto()
+    NEXTN_SHARED_HEAD_HEAD = auto()
+    NEXTN_SHARED_HEAD_NORM = auto()
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -678,6 +687,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.DEEPSEEK2:        "deepseek2",
     MODEL_ARCH.CHATGLM:          "chatglm",
     MODEL_ARCH.GLM4:             "glm4",
+    MODEL_ARCH.GLM4_MOE:         "glm4moe",
     MODEL_ARCH.BITNET:           "bitnet",
     MODEL_ARCH.T5:               "t5",
     MODEL_ARCH.T5ENCODER:        "t5encoder",
@@ -936,6 +946,13 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.A_MMPROJ_FC:               "mm.a.fc",
     MODEL_TENSOR.A_MM_NORM_PRE:             "mm.a.norm_pre",
     MODEL_TENSOR.A_MM_NORM_MID:             "mm.a.norm_mid",
+    # NextN/MTP
+    MODEL_TENSOR.NEXTN_EH_PROJ:             "blk.{bid}.nextn.eh_proj",
+    MODEL_TENSOR.NEXTN_EMBED_TOKENS:        "blk.{bid}.nextn.embed_tokens",
+    MODEL_TENSOR.NEXTN_ENORM:               "blk.{bid}.nextn.enorm",
+    MODEL_TENSOR.NEXTN_HNORM:               "blk.{bid}.nextn.hnorm",
+    MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD:    "blk.{bid}.nextn.shared_head_head",
+    MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM:    "blk.{bid}.nextn.shared_head_norm",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -2124,6 +2141,37 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ATTN_POST_NORM,
         MODEL_TENSOR.FFN_POST_NORM,
     ],
+    MODEL_ARCH.GLM4_MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        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_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        # NextN/MTP tensors - preserved but unused
+        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.BITNET: [
         MODEL_TENSOR.ATTN_Q,
         MODEL_TENSOR.ATTN_K,

gguf-py/gguf/gguf_writer.py

@@ -753,6 +753,9 @@ class GGUFWriter:
     def add_moe_every_n_layers(self, value: int) -> None:
         self.add_uint32(Keys.LLM.MOE_EVERY_N_LAYERS.format(arch=self.arch), value)
 
+    def add_nextn_predict_layers(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.NEXTN_PREDICT_LAYERS.format(arch=self.arch), count)
+
     def add_swin_norm(self, value: bool) -> None:
         self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value)
 

gguf-py/gguf/scripts/gguf_new_metadata.py

@@ -111,6 +111,7 @@ def main() -> None:
     parser.add_argument("--general-description",                       type=str,  help="The models general.description", metavar='"Description ..."')
     parser.add_argument("--chat-template",                             type=str,  help="Chat template string (or JSON string containing templates)", metavar='"{% ... %} ..."')
     parser.add_argument("--chat-template-config",                      type=Path, help="Config file containing chat template(s)", metavar='tokenizer_config.json')
+    parser.add_argument("--chat-template-file",                        type=Path, help="Jinja file containing chat template", metavar='chat_template.jinja')
     parser.add_argument("--pre-tokenizer",                             type=str,  help="The models tokenizer.ggml.pre", metavar='"pre tokenizer"')
     parser.add_argument("--remove-metadata",      action="append",     type=str,  help="Remove metadata (by key name) from output model", metavar='general.url')
     parser.add_argument("--special-token",        action="append",     type=str,  help="Special token by value", nargs=2, metavar=(' | '.join(token_names.keys()), '"<token>"'))
@@ -134,12 +135,17 @@ def main() -> None:
         new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, json.loads(args.chat_template) if args.chat_template.startswith('[') else args.chat_template)
 
     if args.chat_template_config:
-        with open(args.chat_template_config, 'r') as fp:
+        with open(args.chat_template_config, 'r', encoding='utf-8') as fp:
             config = json.load(fp)
             template = config.get('chat_template')
             if template:
                 new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template)
 
+    if args.chat_template_file:
+        with open(args.chat_template_file, 'r', encoding='utf-8') as fp:
+            template = fp.read()
+            new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template)
+
     if args.pre_tokenizer:
         new_metadata[gguf.Keys.Tokenizer.PRE] = MetadataDetails(gguf.GGUFValueType.STRING, args.pre_tokenizer)
 

gguf-py/gguf/tensor_mapping.py

@@ -1369,6 +1369,31 @@ class TensorNameMap:
         MODEL_TENSOR.A_MM_NORM_MID: (
             "audio.multi_modal_projector.ln_mid", # ultravox
         ),
+
+        # NextN/MTP tensors for GLM4_MOE
+        MODEL_TENSOR.NEXTN_EH_PROJ: (
+            "model.layers.{bid}.eh_proj",
+        ),
+
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS: (
+            "model.layers.{bid}.embed_tokens",
+        ),
+
+        MODEL_TENSOR.NEXTN_ENORM: (
+            "model.layers.{bid}.enorm",
+        ),
+
+        MODEL_TENSOR.NEXTN_HNORM: (
+            "model.layers.{bid}.hnorm",
+        ),
+
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD: (
+            "model.layers.{bid}.shared_head.head",
+        ),
+
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM: (
+            "model.layers.{bid}.shared_head.norm",
+        ),
     }
 
     # architecture-specific block mappings

gguf-py/gguf/vocab.py

@@ -312,7 +312,11 @@ class SpecialVocab:
         with open(config_file, encoding = 'utf-8') as f:
             config = json.load(f)
         for typ in self.special_token_types:
-            self._set_special_token(typ, config.get(f'{typ}_token_id'))
+            token_id = config.get(f'{typ}_token_id')
+            # If not found at root, check in text_config (for multimodal models like Kimi-VL)
+            if token_id is None and 'text_config' in config:
+                token_id = config['text_config'].get(f'{typ}_token_id')
+            self._set_special_token(typ, token_id)
         return True
 
 

models/templates/README.md

@@ -21,4 +21,5 @@ These templates can be updated with the following commands:
 ./scripts/get_chat_template.py Qwen/Qwen2.5-7B-Instruct                      > models/templates/Qwen-Qwen2.5-7B-Instruct.jinja
 ./scripts/get_chat_template.py Qwen/QwQ-32B                                  > models/templates/Qwen-QwQ-32B.jinja
 ./scripts/get_chat_template.py Qwen/Qwen3-0.6B                               > models/templates/Qwen-Qwen3-0.6B.jinja
-```
+./scripts/get_chat_template.py zai-org/GLM-4.5                               > models/templates/zai-org-GLM-4.5.jinja
+```

src/llama-arch.cpp

@@ -62,6 +62,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DEEPSEEK2,        "deepseek2"        },
     { LLM_ARCH_CHATGLM,          "chatglm"          },
     { LLM_ARCH_GLM4,             "glm4"             },
+    { LLM_ARCH_GLM4_MOE,         "glm4moe"          },
     { LLM_ARCH_BITNET,           "bitnet"           },
     { LLM_ARCH_T5,               "t5"               },
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
@@ -127,6 +128,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
     { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
+    { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -1391,6 +1393,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
         },
     },
+    {
+        LLM_ARCH_GLM4_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,     "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+            // NextN/MTP tensors - preserved but unused (in final layer, dynamic layer number)
+            { LLM_TENSOR_NEXTN_EH_PROJ,      "blk.%d.nextn.eh_proj" },
+            { LLM_TENSOR_NEXTN_EMBED_TOKENS, "blk.%d.nextn.embed_tokens" },
+            { LLM_TENSOR_NEXTN_ENORM,        "blk.%d.nextn.enorm" },
+            { LLM_TENSOR_NEXTN_HNORM,        "blk.%d.nextn.hnorm" },
+            { LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "blk.%d.nextn.shared_head_head" },
+            { LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "blk.%d.nextn.shared_head_norm" },
+        },
+    },
     {
         LLM_ARCH_BITNET,
         {
@@ -2181,6 +2217,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SHORTCONV_CONV,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SHORTCONV_INPROJ,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SHORTCONV_OUTPROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    // NextN/MTP tensors are currently ignored (reserved for future MTP support)
+    // These tensors only exist in the last layer(s) and are treated as output tensors
+    {LLM_TENSOR_NEXTN_EH_PROJ,              {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_NEXTN_EMBED_TOKENS,         {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_NEXTN_ENORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_NEXTN_HNORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
+    {LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
 };
 
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}

src/llama-arch.h

@@ -66,6 +66,7 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
     LLM_ARCH_GLM4,
+    LLM_ARCH_GLM4_MOE,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
@@ -131,6 +132,7 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,
     LLM_KV_MOE_EVERY_N_LAYERS,
+    LLM_KV_NEXTN_PREDICT_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
@@ -409,6 +411,12 @@ enum llm_tensor {
     LLM_TENSOR_SHORTCONV_CONV,
     LLM_TENSOR_SHORTCONV_INPROJ,
     LLM_TENSOR_SHORTCONV_OUTPROJ,
+    LLM_TENSOR_NEXTN_EH_PROJ,
+    LLM_TENSOR_NEXTN_EMBED_TOKENS,
+    LLM_TENSOR_NEXTN_ENORM,
+    LLM_TENSOR_NEXTN_HNORM,
+    LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
+    LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
 };
 
 enum llm_tensor_layer {

src/llama-graph.cpp

@@ -749,8 +749,8 @@ ggml_tensor * llm_graph_context::build_ffn(
 
     if (down) {
         cur = build_lora_mm(down, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
+            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
             ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
         }
     }
@@ -1391,8 +1391,8 @@ ggml_tensor * llm_graph_context::build_attn(
 
     if (wo) {
         cur = build_lora_mm(wo, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
+            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
             ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
         }
     }

src/llama-hparams.h

@@ -73,6 +73,7 @@ struct llama_hparams {
     bool     expert_weights_norm  = false;
     uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
     uint32_t moe_every_n_layers   = 0;
+    uint32_t nextn_predict_layers = 0;
 
     float f_norm_eps;
     float f_norm_rms_eps;

src/llama-kv-cache-unified.cpp

@@ -39,6 +39,10 @@ llama_kv_cache_unified::llama_kv_cache_unified(
     if (model.arch == LLM_ARCH_GEMMA3N) {
         n_layer_cache = 20;
     }
+    if (model.arch == LLM_ARCH_GLM4_MOE) {
+        // GLM-4.5: Only process up to last layer, skip final NextN layer
+        n_layer_cache = hparams.n_layer - hparams.nextn_predict_layers;
+    }
 
     // create a context for each buffer type
     std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;

src/llama-memory-hybrid.cpp

@@ -25,6 +25,7 @@ llama_memory_hybrid::llama_memory_hybrid(
                          /* common */
              uint32_t    n_seq_max,
                  bool    offload,
+                 bool    unified,
                          /* layer filters */
       layer_filter_cb && filter_attn,
       layer_filter_cb && filter_recr) :
@@ -38,7 +39,7 @@ llama_memory_hybrid::llama_memory_hybrid(
         type_v,
         v_trans,
         offload,
-        1,
+        unified,
         kv_size,
         n_seq_max,
         n_pad,

src/llama-memory-hybrid.h

@@ -39,6 +39,7 @@ public:
                              /* common */
                  uint32_t    n_seq_max,
                      bool    offload,
+                     bool    unified,
                              /* layer filters */
           layer_filter_cb && filter_attn = nullptr,
           layer_filter_cb && filter_recr = nullptr);

src/llama-model-loader.h

@@ -58,8 +58,9 @@ struct llama_model_loader {
         }
     };
 
-    static const int TENSOR_NOT_REQUIRED = 1;
-    static const int TENSOR_DUPLICATED   = 2;
+    static const int TENSOR_NOT_REQUIRED = 1 << 0;
+    static const int TENSOR_DUPLICATED   = 1 << 1;
+    static const int TENSOR_SKIP         = 1 << 2;
 
     int n_kv      = 0;
     int n_tensors = 0;

src/llama-model.cpp

@@ -109,8 +109,10 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_A13B:          return "A13B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_106B_A12B:     return "106B.A12B";
         case LLM_TYPE_235B_A22B:     return "235B.A22B";
         case LLM_TYPE_300B_A47B:     return "300B.A47B";
+        case LLM_TYPE_355B_A32B:     return "355B.A32B";
         case LLM_TYPE_E2B:           return "E2B";
         case LLM_TYPE_E4B:           return "E4B";
         default:                     return "?B";
@@ -1434,6 +1436,34 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GLM4_MOE:
+            {
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                // MoE parameters
+                ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);
+                ml.get_key(LLM_KV_EXPERT_USED_COUNT,           hparams.n_expert_used);
+                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);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+
+                // Expert gating function (GLM-4.5 uses sigmoid)
+                ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+                if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+                    hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+                }
+
+                // NextN/MTP parameters
+                ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,        hparams.nextn_predict_layers, false);
+
+                switch (hparams.n_layer) {
+                    case 47: type = LLM_TYPE_106B_A12B; break; // GLM-4.5-Air (46 layers + 1 NextN layer)
+                    case 93: type = LLM_TYPE_355B_A32B; break; // GLM-4.5 (92 layers + 1 NextN layer)
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_BITNET:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1949,6 +1979,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
     const auto TENSOR_DUPLICATED   = llama_model_loader::TENSOR_DUPLICATED;
     const auto TENSOR_NOT_REQUIRED = llama_model_loader::TENSOR_NOT_REQUIRED;
+    const auto TENSOR_SKIP         = llama_model_loader::TENSOR_SKIP;
 
     // create tensors for the weights
     {
@@ -2004,7 +2035,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             }
 
             // skip unused tensors
-            if (info.op == GGML_OP_NONE) {
+            if (info.op == GGML_OP_NONE || flags & TENSOR_SKIP) {
                 const size_t nbytes = ggml_nbytes(t_meta);
                 LLAMA_LOG_WARN("model has unused tensor %s (size = %zu bytes) -- ignoring\n", tn.str().c_str(), nbytes);
 
@@ -4427,6 +4458,105 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_GLM4_MOE:
+                {
+                    const int64_t n_expert        = hparams.n_expert;
+                    const int64_t n_expert_used   = hparams.n_expert_used;
+                    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+                    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");
+
+                    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);
+                    }
+
+                    // Load ALL tensors including NextN layer to satisfy total tensor count
+                    // but only PROCESS up to last layer (skipping final NextN layer) in forward pass
+                    for (int i = 0; i < n_layer; ++i) {
+                        int flags = 0;
+                        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+                            // skip all tensors in the NextN layers
+                            flags |= TENSOR_SKIP;
+                        }
+
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
+
+                        // GLM-style attention with bias terms
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, flags);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, flags);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, flags);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), { n_embd_head_k * n_head }, flags);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), { n_embd_k_gqa }, flags);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), { n_embd_v_gqa }, flags);
+
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags);
+
+                        // K/Q norm tensors (optional for GLM-4.5 355B variant)
+                        layer.attn_q_norm = create_tensor(
+                            tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags);
+                        layer.attn_k_norm = create_tensor(
+                            tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags);
+
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, flags);
+
+                        // Check if this layer uses MoE or dense FFN based on n_layer_dense_lead
+                        // GLM 4.5 uses hybrid architecture: layer 0 is dense, layers 1+ are MoE
+                        const bool use_moe = (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead);
+
+                        if (use_moe) {
+                            // MoE layers
+                            layer.ffn_gate_inp =
+                                create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, flags);
+                            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;
+
+                            layer.ffn_gate_exps = create_tensor(
+                                tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, 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);
+                            layer.ffn_up_exps = create_tensor(
+                                tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags);
+
+                            // Shared expert
+                            if (n_expert_shared > 0) {
+                                const int64_t n_ff_shexp = n_ff_exp * 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);
+                            }
+                        } else {
+                            // Dense layers (first k layers) - GLM uses separate gate/up projections
+                            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);
+                        }
+
+                        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+                        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+                            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
+                            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, 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.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags);
+                            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags);
+                        }
+                    }
+                }
+                break;
             case LLM_ARCH_NEMOTRON:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -13564,6 +13694,169 @@ struct llm_build_glm4 : public llm_graph_context {
     }
 };
 
+struct llm_build_glm4_moe : public llm_graph_context {
+    llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        // Only process up to last layer (skip final NextN layer)
+        // Final layer tensors are loaded but not processed in forward pass
+        const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+        for (int il = 0; il < n_transformer_layers; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // Pre-attention norm
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                cb(Kcur, "Kcur", il);
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                // Apply Q/K norm if available (GLM-4.5 355B variant)
+                if (model.layers[il].attn_q_norm) {
+                    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                    cb(Qcur, "Qcur_normed", il);
+                }
+                if (model.layers[il].attn_k_norm) {
+                    Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                    cb(Kcur, "Kcur_normed", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_transformer_layers - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // Post-attention norm
+            cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "post_attn_norm", il);
+
+            // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
+            if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+                // Dense FFN layer
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                // MoE layer with shared experts
+                const int64_t n_expert      = hparams.n_expert;
+                const int64_t n_expert_used = hparams.n_expert_used;
+
+                // Process routed experts using existing MoE infrastructure
+                ggml_tensor * routed_out = build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        model.layers[il].ffn_exp_probs_b,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, hparams.expert_weights_norm,
+                        true, hparams.expert_weights_scale,
+                        (llama_expert_gating_func_type) hparams.expert_gating_func,
+                        il);
+                cb(routed_out, "ffn_moe_out", il);
+
+                // Process shared expert on original input
+                ggml_tensor * shared_out = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(shared_out, "ffn_shexp_out", il);
+
+                // Final output: routed_output + shared_output
+                cur = ggml_add(ctx0, routed_out, shared_out);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 struct llm_build_nemotron : public llm_graph_context {
     llm_build_nemotron(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -17598,6 +17891,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
                         /* n_seq_max         */ cparams.n_seq_max,
                         /* offload           */ cparams.offload_kqv,
+                        /* unified           */ cparams.kv_unified,
                         /* filter_attn       */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr,
                         /* filter_recr       */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr);
                 } else {
@@ -17876,6 +18170,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_glm4>(*this, params);
             } break;
+        case LLM_ARCH_GLM4_MOE:
+            {
+                llm = std::make_unique<llm_build_glm4_moe>(*this, params);
+            } break;
         case LLM_ARCH_BITNET:
             {
                 llm = std::make_unique<llm_build_bitnet>(*this, params);
@@ -18207,6 +18505,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_HUNYUAN_DENSE:
         case LLM_ARCH_LFM2:
         case LLM_ARCH_SMALLTHINKER:
+        case LLM_ARCH_GLM4_MOE:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:

src/llama-model.h

@@ -101,8 +101,10 @@ enum llm_type {
     LLM_TYPE_A13B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
+    LLM_TYPE_106B_A12B, // GLM-4.5-Air
     LLM_TYPE_235B_A22B,
     LLM_TYPE_300B_A47B, // Ernie MoE big
+    LLM_TYPE_355B_A32B, // GLM-4.5
     LLM_TYPE_E2B,
     LLM_TYPE_E4B,
 };
@@ -166,6 +168,15 @@ struct llama_layer_shortconv {
     struct ggml_tensor * out_proj = nullptr;
 };
 
+struct llama_layer_nextn {
+    struct ggml_tensor * eh_proj          = nullptr;
+    struct ggml_tensor * embed_tokens     = nullptr;
+    struct ggml_tensor * enorm            = nullptr;
+    struct ggml_tensor * hnorm            = nullptr;
+    struct ggml_tensor * shared_head_head = nullptr;
+    struct ggml_tensor * shared_head_norm = nullptr;
+};
+
 struct llama_layer {
     // normalization
     struct ggml_tensor * attn_norm       = nullptr;
@@ -354,6 +365,8 @@ struct llama_layer {
     struct llama_layer_convnext convnext;
 
     struct llama_layer_shortconv shortconv;
+
+    struct llama_layer_nextn nextn;
 };
 
 struct llama_model {

src/llama-vocab.cpp

@@ -1856,7 +1856,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "gigachat"   ||
                     tokenizer_pre == "jina-v2-es" ||
                     tokenizer_pre == "jina-v2-de" ||
-                    tokenizer_pre == "a.x-4.0") {
+                    tokenizer_pre == "a.x-4.0" ||
+                    tokenizer_pre == "mellum") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                     tokenizer_pre == "jina-v1-en" ||
@@ -2190,6 +2191,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<｜fim▁begin｜>" // DeepSeek
                         || t.first == "<PRE>"
                         || t.first == "▁<PRE>"          // CodeLlama
+                        || t.first == "<|code_prefix|>" // GLM-4.5
                         ) {
                     special_fim_pre_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2209,6 +2211,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<｜fim▁hole｜>" // DeepSeek
                         || t.first == "<SUF>"
                         || t.first == "▁<SUF>"         // CodeLlama
+                        || t.first == "<|code_suffix|>" // GLM-4.5
                         ) {
                     special_fim_suf_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2228,6 +2231,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<｜fim▁end｜>"  // DeepSeek
                         || t.first == "<MID>"
                         || t.first == "▁<MID>"         // CodeLlama
+                        || t.first == "<|code_middle|>" // GLM-4.5
                         ) {
                     special_fim_mid_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {

tools/imatrix/README.md

@@ -7,7 +7,7 @@ More information is available in <https://github.com/ggml-org/llama.cpp/pull/486
 
 ```
 ./llama-imatrix \
-    -m model.gguf -f some-text.txt [-o imatrix.gguf] [--no-ppl] \
+    -m model.gguf -f some-text.txt [-o imatrix.gguf] [--output-format {gguf,dat}] [--no-ppl] \
     [--process-output] [--chunk 123] [--save-frequency 0] [--output-frequency 10] \
     [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] [--parse-special] \
     [--show-statistics] [...]
@@ -20,6 +20,7 @@ The parameters in square brackets are optional and have the following meaning:
 * `-lv | --verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
 * `-o | --output-file` specifies the name of the file where the computed data will be stored. If missing `imatrix.gguf` is used.
 * `-ofreq | --output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
+* `--output-format` specifies the output format of the generated imatrix file. Either "gguf", or "dat" (the legacy format). Defaults to "gguf".
 * `--save-frequency` specifies how often to save a copy of the imatrix in a separate file. Default is 0 (i.e., never)
 * `--process-output` specifies if data will be collected for the `output.weight` tensor. Typically, it is better not to utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
 * `--in-file` one or more existing imatrix files to load and combine. Useful for merging files from multiple runs/datasets.
@@ -45,14 +46,19 @@ Recent versions of `llama-imatrix` store data in GGUF format by default. For the
 
 ```bash
 # generate and save the imatrix using legacy format
-./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt -o imatrix-legcy-format.dat -ngl 99
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt --output-format dat -o imatrix-legcy-format.dat -ngl 99
 ```
 
 ```bash
-# covert legacy (binary) imatrix format to new (GGUF) format
+# convert legacy (binary) imatrix format to new (GGUF) format
 ./llama-imatrix --in-file imatrix-legacy-format.dat -o imatrix-new-format.gguf
 ```
 
+```bash
+# convert new (GGUF) imatrix format to legacy (binary) format
+./llama-imatrix --in-file imatrix-new-format.gguf --output-format dat -o imatrix-legacy-format.dat
+```
+
 ```bash
 # combine existing imatrices
 ./llama-imatrix --in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf -o imatrix-combined.gguf

tools/imatrix/imatrix.cpp

@@ -26,7 +26,7 @@
 static void print_usage(int, char ** argv) {
     LOG("\nexample usage:\n");
     LOG("\n    %s \\\n"
-            "       -m model.gguf -f some-text.txt [-o imatrix.gguf] [--no-ppl] \\\n"
+            "       -m model.gguf -f some-text.txt [-o imatrix.gguf] [--output-format {gguf,dat}] [--no-ppl] \\\n"
             "       [--process-output] [--chunk 123] [--save-frequency 0] [--output-frequency 10] \\\n"
             "       [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] [--parse-special] \\\n"
             "       [--show-statistics] [...]\n" , argv[0]);
@@ -250,13 +250,6 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
     const char * data = is_host ? (const char *) src1->data : m_src1_data.data();
     GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
 
-    // TODO: 4d? (is that even used in practice?)
-    // the extra dimension would need to be stored somewhere to be reflected in the imatrix file
-    if (ggml_nrows(src1) != src1->ne[1] * src1->ne[2]) {
-        LOG_ERR("%s: tensor has more than 3 dimensions: %s", __func__, wname.c_str());
-        GGML_ASSERT(false);
-    }
-
     // this has been adapted to the new format of storing merged experts in a single 3d tensor
     // ref: https://github.com/ggml-org/llama.cpp/pull/6387
     if (t->op == GGML_OP_MUL_MAT_ID) {
@@ -272,6 +265,12 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
 
         GGML_ASSERT(ids->ne[1] == src1->ne[2]);
 
+        // the extra dimension would need to be stored somewhere to be reflected in the imatrix file
+        if (ggml_nrows(src1) != src1->ne[1] * src1->ne[2]) {
+            LOG_ERR("%s: tensor has more than 3 dimensions: %s", __func__, wname.c_str());
+            GGML_ASSERT(false);
+        }
+
         m_ids.resize(ggml_nbytes(ids));
         ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids));
 
@@ -335,29 +334,40 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         }
     } else {
         auto & e = m_stats[wname];
-        const int64_t n_mat = src1->ne[2] * src1->ne[3];
+        const int64_t n_mat = src0->ne[2] * src0->ne[3];
 
+        // use a single count per dense tensor
+        // (necessary when merging older GGUF-imatrix files with 3d tensors)
+        if (e.counts.size() > 1) {
+            bool all_equal = true;
+            for (size_t i = 1; i < e.counts.size(); ++i) {
+                if (e.counts[0] != e.counts[i]) {
+                    all_equal = false;
+                    break;
+                }
+            }
+            if (all_equal) {
+                e.counts.resize(1);
+            }
+        }
         if (e.values.empty()) {
             e.values.resize(src1->ne[0] * n_mat, 0);
-            e.counts.resize(n_mat, 0);
+            e.counts.resize(1, 0);
         }
         else if (e.values.size() != (size_t)(src1->ne[0] * n_mat)) {
             LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)(src1->ne[0] * n_mat));
             exit(1); //GGML_ABORT("fatal error");
         }
-        else if (e.counts.size() != (size_t)n_mat) {
-            LOG_ERR("%s: inconsistent expert count for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.counts.size(), (int)n_mat);
-            exit(1); //GGML_ABORT("fatal error");
-        }
         LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->ne[2], (int)src1->type);
+
         for (int64_t i3 = 0; i3 < src1->ne[3]; ++i3) {
             for (int64_t i2 = 0; i2 < src1->ne[2]; ++i2) {
-                const int64_t mat_id = i3 * src1->ne[2] + i2;
+                // handle 3D+ tensors, but flatten 3D+ activations when model tensor is 2D
+                const int64_t mat_id = (i3 % src0->ne[3]) * src0->ne[2] + (i2 % src0->ne[2]);
                 const int64_t mat_start = mat_id * src1->ne[0];
 
                 for (int64_t row = 0; row < src1->ne[1]; ++row) {
-                    const float * x = (const float *) (data + row * src1->nb[1] + i2 * src1->nb[2] + i3 * src1->ne[3]);
-                    e.counts[mat_id]++;
+                    const float * x = (const float *) (data + row * src1->nb[1] + i2 * src1->nb[2] + i3 * src1->nb[3]);
                     for (int64_t j = 0; j < src1->ne[0]; ++j) {
                         e.values[mat_start + j] += x[j] * x[j];
                         if (!std::isfinite((float)e.values[j])) {
@@ -366,16 +376,20 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
                         }
                     }
                 }
-                const int32_t n_chunk = e.counts[mat_id] / chunk_size;
-                if (n_chunk > m_last_chunk) {
-                    const int32_t chunk_step = n_chunk - m_last_chunk;
-                    m_last_chunk = n_chunk;
-                    if ((m_last_chunk % m_params.n_out_freq) / chunk_step == 0) {
-                        save_imatrix();
-                    }
-                    if (m_params.n_save_freq > 0 && (m_last_chunk % m_params.n_save_freq) / chunk_step == 0) {
-                        save_imatrix(m_last_chunk);
-                    }
+            }
+        }
+        // only 1 count in practice, except when a tensor is used for both MUL_MAT_ID and MUL_MAT
+        for (size_t i = 0; i < e.counts.size(); ++i) {
+            e.counts[i] += ggml_nrows(src1) / n_mat;
+            const int32_t n_chunk = e.counts[i] / chunk_size;
+            if (n_chunk > m_last_chunk) {
+                const int32_t chunk_step = n_chunk - m_last_chunk;
+                m_last_chunk = n_chunk;
+                if ((m_last_chunk % m_params.n_out_freq) / chunk_step == 0) {
+                    save_imatrix();
+                }
+                if (m_params.n_save_freq > 0 && (m_last_chunk % m_params.n_save_freq) / chunk_step == 0) {
+                    save_imatrix(m_last_chunk);
                 }
             }
         }
@@ -492,13 +506,17 @@ void IMatrixCollector::save_imatrix_legacy(int32_t ncall) const {
 
 void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
     auto fname = m_params.out_file;
+    int8_t use_legacy_format = m_params.imat_dat;
 
-    // TODO: use the new format in more cases
-    if (!string_ends_with(fname, ".gguf")) {
-        LOG_WRN("\n%s: saving to legacy imatrix format because output suffix is not .gguf\n", __func__);
+    if (use_legacy_format > 0) {
         this->save_imatrix_legacy(n_chunk);
         return;
     }
+    // only warn when `--output-format gguf` is not specified
+    if (use_legacy_format == 0 && !string_ends_with(fname, ".gguf")) {
+        LOG_WRN("\n%s: saving imatrix using GGUF format with a different suffix than .gguf\n", __func__);
+        LOG_WRN("%s: if you want the previous imatrix format, use --output-format dat\n", __func__);
+    }
 
     if (n_chunk > 0) {
         fname += ".at_";

tools/quantize/quantize.cpp

@@ -611,7 +611,7 @@ int main(int argc, char ** argv) {
             return 1;
         }
         if (!try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
-            fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[3]);
+            fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[arg_idx]);
             return 1;
         }
         if (ftype_str == "COPY") {