CANN: refactor mask handling and improve performance in FA (#15561)

* CANN(flash-attn): refactor mask handling and improve performance

1. Refactored the mask computation in Flash Attention, unified the logic without separating prefill and decode.
2. Optimized performance in non-alibi scenarios by reducing one repeat operation.
3. Updated operator management to explicitly mark unsupported cases on 310P devices and when dim is not divisible by 16.

Signed-off-by: noemotiovon <757486878@qq.com>

* [CANN]: fix review

Signed-off-by: noemotiovon <757486878@qq.com>

* [CANN]: Optimization FA BNSD to BSND

Signed-off-by: noemotiovon <757486878@qq.com>

---------

Signed-off-by: noemotiovon <757486878@qq.com>

.devops/vulkan.Dockerfile

@@ -2,14 +2,30 @@ ARG UBUNTU_VERSION=24.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
-# Install build tools
-RUN apt update && apt install -y git build-essential cmake wget
+# Ref: https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html
 
-# Install Vulkan SDK and cURL
-RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
-    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-noble.list https://packages.lunarg.com/vulkan/lunarg-vulkan-noble.list && \
-    apt update -y && \
-    apt-get install -y vulkan-sdk libcurl4-openssl-dev curl
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget xz-utils
+
+# Install Vulkan SDK
+ARG VULKAN_VERSION=1.4.321.1
+RUN ARCH=$(uname -m) && \
+    wget -qO /tmp/vulkan-sdk.tar.xz https://sdk.lunarg.com/sdk/download/${VULKAN_VERSION}/linux/vulkan-sdk-linux-${ARCH}-${VULKAN_VERSION}.tar.xz && \
+    mkdir -p /opt/vulkan && \
+    tar -xf /tmp/vulkan-sdk.tar.xz -C /tmp --strip-components=1 && \
+    mv /tmp/${ARCH}/* /opt/vulkan/ && \
+    rm -rf /tmp/*
+
+# Install cURL and Vulkan SDK dependencies
+RUN apt install -y libcurl4-openssl-dev curl \
+    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev
+
+# Set environment variables
+ENV VULKAN_SDK=/opt/vulkan
+ENV PATH=$VULKAN_SDK/bin:$PATH
+ENV LD_LIBRARY_PATH=$VULKAN_SDK/lib:$LD_LIBRARY_PATH
+ENV CMAKE_PREFIX_PATH=$VULKAN_SDK:$CMAKE_PREFIX_PATH
+ENV PKG_CONFIG_PATH=$VULKAN_SDK/lib/pkgconfig:$PKG_CONFIG_PATH
 
 # Build it
 WORKDIR /app

.github/workflows/build-riscv-native.yml

@@ -1,10 +1,11 @@
 name: Build on RISCV Linux Machine by Cloud-V
 on:
+  pull_request:
   workflow_dispatch:
   workflow_call:
 
 jobs:
-  bianbu-riscv64-native: # Bianbu 2.2
+  debian-13-riscv64-native: # Bianbu 2.2
     runs-on: self-hosted
 
     steps:
@@ -20,24 +21,40 @@ jobs:
                   build-essential \
                   gcc-14-riscv64-linux-gnu \
                   g++-14-riscv64-linux-gnu \
+                  ccache \
                   cmake
 
+      - name: Setup ccache
+        run: |
+          mkdir -p $HOME/.ccache
+          ccache -M 5G -d $HOME/.ccache
+          export CCACHE_LOGFILE=/home/runneruser/ccache_debug/ccache.log
+          export CCACHE_DEBUGDIR="/home/runneruser/ccache_debug"
+          echo "$GITHUB_WORKSPACE"
+          echo "CCACHE_LOGFILE=$CCACHE_LOGFILE" >> $GITHUB_ENV
+          echo "CCACHE_DEBUGDIR=$CCACHE_DEBUGDIR" >> $GITHUB_ENV
+          echo "CCACHE_BASEDIR=$GITHUB_WORKSPACE" >> $GITHUB_ENV
+          echo "CCACHE_DIR=$HOME/.ccache" >> $GITHUB_ENV
+
       - name: Build
         run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+          cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DCMAKE_SYSTEM_NAME=Linux \
+            -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+            -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+            -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+            -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+            -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)

README.md

@@ -151,6 +151,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 - [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge)
 - [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
+- [x] [LFM2-VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa)
 
 </details>
 

ci/run.sh

@@ -106,7 +106,7 @@ function gg_wget {
     cd $out
 
     # should not re-download if file is the same
-    wget -nv -N $url
+    wget -nv -c -N $url
 
     cd $cwd
 }

common/arg.cpp

@@ -1106,7 +1106,7 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
     printf("\"\n\n");
 
     printf("    case \"$prev\" in\n");
-    printf("        --model)\n");
+    printf("        --model|-m)\n");
     printf("            COMPREPLY=( $(compgen -f -X '!*.gguf' -- \"$cur\") $(compgen -d -- \"$cur\") )\n");
     printf("            return 0\n");
     printf("            ;;\n");
@@ -1755,7 +1755,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.warmup = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -2254,9 +2254,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"-dt", "--defrag-thold"}, "N",
-        string_format("KV cache defragmentation threshold (default: %.1f, < 0 - disabled)", (double)params.defrag_thold),
+        string_format("KV cache defragmentation threshold (DEPRECATED)"),
         [](common_params & params, const std::string & value) {
-            params.defrag_thold = std::stof(value);
+            GGML_UNUSED(params);
+            GGML_UNUSED(value);
+            LOG_WRN("DEPRECATED: --defrag-thold is deprecated and no longer necessary to specify\n");
         }
     ).set_env("LLAMA_ARG_DEFRAG_THOLD"));
     add_opt(common_arg(

common/chat.cpp

@@ -1361,6 +1361,26 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
         "<|end|>",
     };
 
+    if (!inputs.json_schema.is_null()) {
+        data.grammar_lazy = false;
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            auto schema = inputs.json_schema;
+            builder.resolve_refs(schema);
+
+            auto not_end = builder.add_rule("not-end",
+                "[^<] | \"<\" [^|] | \"<|\" [^e] | \"<|e\" [^n] | \"<|en\" [^d] | \"<|end\" [^|] | \"<|end|\" [^>]");
+            auto analysis = builder.add_rule("analysis",
+                "\"<|channel|>analysis<|message|>\" ( " + not_end + " )* \"<|end|>\"");
+            auto constraint = builder.add_rule("constraint", "\"<|constrain|>\"? [a-zA-Z0-9_-]+");
+            auto final = builder.add_rule("final",
+                "\"<|channel|>final\" ( \" \" " + constraint + " )? \"<|message|>\" " +
+                builder.add_schema("response", schema)
+            );
+
+            builder.add_rule("root", "( " + analysis + " \"<|start|>assistant\" )? " + final);
+        });
+    }
+
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
         data.grammar = build_grammar([&](const common_grammar_builder & builder) {
@@ -2121,7 +2141,7 @@ static common_chat_params common_chat_templates_apply_jinja(
     }
 
     // GPT-OSS
-    if (src.find("<|channel|>") != std::string::npos && params.json_schema.is_null()) {
+    if (src.find("<|channel|>") != std::string::npos) {
         return common_chat_params_init_gpt_oss(tmpl, params);
     }
 

common/common.cpp

@@ -1152,7 +1152,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
     cparams.pooling_type      = params.pooling_type;
     cparams.attention_type    = params.attention_type;
-    cparams.defrag_thold      = params.defrag_thold;
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;

common/common.h

@@ -288,7 +288,6 @@ struct common_params {
     float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
     float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
-    float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
 
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading

convert_hf_to_gguf.py

@@ -1216,6 +1216,55 @@ class TextModel(ModelBase):
                 raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
             self.gguf_writer.add_pooling_type(pooling_type)
 
+    def _set_vocab_interns1(self):
+        tokens: list[str] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())
+        vocab_size = self.hparams.get("vocab_size", len(vocab))
+        assert max(vocab.values()) < vocab_size
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()
+
+        added_tokens_decoder = tokenizer.added_tokens_decoder
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            else:
+                token: str = reverse_vocab[i]
+                if token in added_vocab:
+                    # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
+                    # To avoid unexpected issues - we make sure to normalize non-normalized tokens
+                    if not added_tokens_decoder[i].normalized:
+                        previous_token = token
+                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
+                        if previous_token != token:
+                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
+
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    toktypes.append(gguf.TokenType.NORMAL)
+                tokens.append(token)
+
+        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._set_special_token("bos", 151643)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
 
 class MmprojModel(ModelBase):
     model_type = ModelType.MMPROJ
@@ -2932,7 +2981,8 @@ class Qwen2Model(TextModel):
         if "language_model." in name:
             name = name.replace("language_model.", "") # for InternVL
         if name.startswith("mlp") or name.startswith("multi_modal_projector") \
-                or name.startswith("vision_model") or name.startswith("audio_tower"):
+                or name.startswith("vision_model") or name.startswith("audio_tower") \
+                or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector"):
             # skip vision and audio tensors
             return []
         yield from super().modify_tensors(data_torch, name, bid)
@@ -3109,7 +3159,7 @@ class LLaDAModel(TextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Ernie4_5_ForCausalLM")
+@ModelBase.register("Ernie4_5_ForCausalLM", "Ernie4_5ForCausalLM")
 class Ernie4_5Model(TextModel):
     model_arch = gguf.MODEL_ARCH.ERNIE4_5
 
@@ -3604,6 +3654,19 @@ class Qwen2MoeModel(TextModel):
 class Qwen3Model(Qwen2Model):
     model_arch = gguf.MODEL_ARCH.QWEN3
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        hparams = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
+        self.origin_hf_arch = hparams.get('architectures', [None])[0]
+
+    def set_vocab(self):
+        # deal with intern-s1-mini
+        if self.origin_hf_arch == 'InternS1ForConditionalGeneration':
+            self._set_vocab_interns1()
+            return
+
+        super().set_vocab()
+
 
 @ModelBase.register("Qwen3MoeForCausalLM")
 class Qwen3MoeModel(Qwen2MoeModel):
@@ -3620,73 +3683,7 @@ class Qwen3MoeModel(Qwen2MoeModel):
             self._set_vocab_interns1()
             return
 
-        try:
-            self._set_vocab_sentencepiece()
-        except FileNotFoundError:
-            self._set_vocab_gpt2()
-
-    def _set_vocab_interns1(self):
-        tokens: list[str] = []
-        toktypes: list[int] = []
-
-        from transformers import AutoTokenizer
-        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
-        vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())
-        vocab_size = self.hparams.get("vocab_size", len(vocab))
-        assert max(vocab.values()) < vocab_size
-
-        tokpre = self.get_vocab_base_pre(tokenizer)
-
-        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab.items()}
-        added_vocab = tokenizer.get_added_vocab()
-
-        added_tokens_decoder = tokenizer.added_tokens_decoder
-
-        for i in range(vocab_size):
-            if i not in reverse_vocab:
-                tokens.append(f"[PAD{i}]")
-                toktypes.append(gguf.TokenType.UNUSED)
-            else:
-                token: str = reverse_vocab[i]
-                if token in added_vocab:
-                    # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
-                    # To avoid unexpected issues - we make sure to normalize non-normalized tokens
-                    if not added_tokens_decoder[i].normalized:
-                        previous_token = token
-                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
-                        if previous_token != token:
-                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
-
-                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
-                        toktypes.append(gguf.TokenType.CONTROL)
-                    else:
-                        toktypes.append(gguf.TokenType.USER_DEFINED)
-                else:
-                    toktypes.append(gguf.TokenType.NORMAL)
-                tokens.append(token)
-
-        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_tokens_map_file = self.dir_model / 'special_tokens_map.json'
-        additional_special_tokens = []
-        if special_tokens_map_file.is_file():
-            with open(special_tokens_map_file, encoding = 'utf-8') as f:
-                additional_special_tokens = json.load(f).get('additional_special_tokens', [])
-        tokenizer_cfg_file = self.dir_model / 'special_tokens_map.json'
-        if tokenizer_cfg_file.is_file():
-            with open(tokenizer_cfg_file, encoding = 'utf-8') as f:
-                added_tokens_decoder = json.load(f).get('added_tokens_decoder', {})
-                token2ids_map = {data['content'] : int(token) for token, data in added_tokens_decoder.items() if data['special']}
-                for token in additional_special_tokens:
-                    if token in token2ids_map:
-                        special_vocab._set_special_token(token, token2ids_map[token])
-        special_vocab._set_special_token('eos', 151645)
-        special_vocab._set_special_token("bos", 151643)
-        special_vocab.add_to_gguf(self.gguf_writer)
+        super().set_vocab()
 
 
 @ModelBase.register("GPT2LMHeadModel")
@@ -5854,6 +5851,11 @@ class OlmoModel(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("SeedOssForCausalLM")
+class SeedOssModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.SEED_OSS
+
+
 @ModelBase.register("Olmo2ForCausalLM")
 class Olmo2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.OLMO2
@@ -6252,9 +6254,11 @@ class DeepseekModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("DeepseekV2ForCausalLM")
-@ModelBase.register("DeepseekV3ForCausalLM")
-@ModelBase.register("KimiVLForConditionalGeneration")
+@ModelBase.register(
+    "DeepseekV2ForCausalLM",
+    "DeepseekV3ForCausalLM",
+    "KimiVLForConditionalGeneration",
+)
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
@@ -8505,6 +8509,43 @@ class PixtralModel(LlavaVisionModel):
             return "mm.2.weight"
         return super().map_tensor_name(name, try_suffixes)
 
+
+@ModelBase.register("KimiVLForConditionalGeneration")
+class KimiVLModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        self.hparams_vision["image_size"] = 64 * 14 # for compatibility
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.KIMIVL)
+        self.gguf_writer.add_vision_use_gelu(True)
+        self.gguf_writer.add_vision_projector_scale_factor(2)
+        # eps is the same as pytorch's default value
+        assert self.hparams_vision is not None
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-5))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        is_vision_tensor = "vision_tower" in name or "multi_modal_projector" in name
+
+        if is_vision_tensor:
+            if "pos_emb.weight" in name:
+                data_torch = data_torch.view(data_torch.shape[0] * data_torch.shape[1], data_torch.shape[2])
+            elif "wqkv" in name:
+                split_dim = 0 if "weight" in name else -1
+                wq, wk, wv = data_torch.chunk(3, dim=split_dim)
+                return [
+                    (self.map_tensor_name(name.replace("wqkv", "wq")), wq),
+                    (self.map_tensor_name(name.replace("wqkv", "wk")), wk),
+                    (self.map_tensor_name(name.replace("wqkv", "wv")), wv)
+                ]
+
+            return [(self.map_tensor_name(name), data_torch)]
+
+        return [] # skip other tensors
+
 ###### CONVERSION LOGIC ######
 
 

docs/build-s390x.md

@@ -265,8 +265,9 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 | BF16       | 🚫          | 🚫   | ❓   | ❓    |
 | Q4_0       | ✅          | ✅   | ❓   | ❓    |
 | Q4_1       | ✅          | ✅   | ❓   | ❓    |
-| Q5_0       | 🚫          | 🚫   | ❓   | ❓    |
-| Q5_1       | 🚫          | 🚫   | ❓   | ❓    |
+| MXFP4      | 🚫          | 🚫   | ❓   | ❓    |
+| Q5_0       | ✅          | ✅   | ❓   | ❓    |
+| Q5_1       | ✅          | ✅   | ❓   | ❓    |
 | Q8_0       | ✅          | ✅   | ❓   | ❓    |
 | Q2_K       | 🚫          | 🚫   | ❓   | ❓    |
 | Q3_K       | ✅          | ✅   | ❓   | ❓    |
@@ -291,4 +292,4 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
 
-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on July 31, 2025.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Aug 22, 2025.

docs/multimodal/minicpmv4.0.md

@@ -6,7 +6,7 @@ Download [MiniCPM-V-4](https://huggingface.co/openbmb/MiniCPM-V-4) PyTorch model
 
 
 ### Build llama.cpp
-Readme modification time: 20250206
+Readme modification time: 20250731
 
 If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
 

docs/multimodal/minicpmv4.5.md

@@ -0,0 +1,47 @@
+## MiniCPM-V 4.5
+
+### Prepare models and code
+
+Download [MiniCPM-V-4_5](https://huggingface.co/openbmb/MiniCPM-V-4_5) PyTorch model from huggingface to "MiniCPM-V-4_5" folder.
+
+
+### Build llama.cpp
+Readme modification time: 20250826
+
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+
+Clone llama.cpp:
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+```
+
+Build llama.cpp using `CMake`:
+```bash
+cmake -B build
+cmake --build build --config Release
+```
+
+
+### Usage of MiniCPM-V 4
+
+Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-4_5-gguf) by us)
+
+```bash
+python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-V-4_5
+python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-4_5 --minicpmv-projector ../MiniCPM-V-4_5/minicpmv.projector --output-dir ../MiniCPM-V-4_5/ --minicpmv_version 6
+python ./convert_hf_to_gguf.py ../MiniCPM-V-4_5/model
+
+# quantize int4 version
+./build/bin/llama-quantize ../MiniCPM-V-4_5/model/ggml-model-f16.gguf ../MiniCPM-V-4_5/model/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+
+Inference on Linux or Mac
+```bash
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4_5/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-4_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-4_5/mmproj-model-f16.gguf
+```

examples/CMakeLists.txt

@@ -34,6 +34,7 @@ else()
     add_subdirectory(gen-docs)
     add_subdirectory(training)
     add_subdirectory(diffusion)
+    add_subdirectory(model-conversion)
     if (NOT GGML_BACKEND_DL)
         add_subdirectory(convert-llama2c-to-ggml)
         # these examples use the backends directly and cannot be built with dynamic loading

examples/llama.vim

@@ -17,7 +17,7 @@
 "
 " start the llama.cpp server with a FIM-compatible model. for example:
 "
-"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
+"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa --ubatch-size 512 --batch-size 1024 --cache-reuse 256
 "
 "   --batch-size [512, model max context]
 "

examples/model-conversion/.gitignore

@@ -0,0 +1,3 @@
+.model_name
+data
+ppl

examples/model-conversion/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-logits)
+add_executable(${TARGET} logits.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/model-conversion/Makefile

@@ -0,0 +1,192 @@
+MAKEFLAGS += --no-print-directory
+
+define validate_model_path
+	@if [ -z "$(MODEL_PATH)" ]; then \
+		echo "Error: MODEL_PATH must be provided either as:"; \
+		echo "  1. Environment variable: export MODEL_PATH=/path/to/model"; \
+		echo "  2. Command line argument: make $(1) MODEL_PATH=/path/to/model"; \
+		exit 1; \
+	fi
+endef
+
+define validate_embedding_model_path
+	@if [ -z "$(EMBEDDING_MODEL_PATH)" ]; then \
+		echo "Error: EMBEDDING_MODEL_PATH must be provided either as:"; \
+		echo "  1. Environment variable: export EMBEDDING_MODEL_PATH=/path/to/model"; \
+		echo "  2. Command line argument: make $(1) EMBEDDING_MODEL_PATH=/path/to/model"; \
+		exit 1; \
+	fi
+endef
+
+define quantize_model
+	@CONVERTED_MODEL="$(1)" QUANTIZED_TYPE="$(QUANTIZED_TYPE)" \
+	TOKEN_EMBD_TYPE="$(TOKEN_EMBD_TYPE)" OUTPUT_TYPE="$(OUTPUT_TYPE)" \
+	./scripts/utils/quantize.sh "$(1)" "$(QUANTIZED_TYPE)" "$(TOKEN_EMBD_TYPE)" "$(OUTPUT_TYPE)"
+	@echo "Export the quantized model path to $(2) variable in your environment"
+endef
+
+###
+### Casual Model targets/recipes
+###
+causal-convert-model-bf16: OUTTYPE=bf16
+causal-convert-model-bf16: causal-convert-model
+
+causal-convert-model:
+	$(call validate_model_path,causal-convert-model)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/causal/convert-model.sh
+
+causal-run-original-model:
+	$(call validate_model_path,causal-run-original-model)
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/causal/run-org-model.py
+
+causal-run-converted-model:
+	@CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/causal/run-converted-model.sh
+
+causal-verify-logits: causal-run-original-model causal-run-converted-model
+	@./scripts/causal/compare-logits.py
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/utils/check-nmse.py -m ${MODEL_PATH}
+
+causal-run-original-embeddings:
+	@./scripts/causal/run-casual-gen-embeddings-org.sh
+
+causal-run-converted-embeddings:
+	@./scripts/causal/run-converted-model-embeddings-logits.sh
+
+causal-verify-embeddings: causal-run-original-embeddings causal-run-converted-embeddings
+	@./scripts/causal/compare-embeddings-logits.sh
+
+causal-inspect-original-model:
+	@./scripts/utils/inspect-org-model.py
+
+causal-inspect-converted-model:
+	@./scripts/utils/inspect-converted-model.sh
+
+causal-start-embedding-server:
+	@./scripts/utils/run-embedding-server.sh ${CONVERTED_MODEL}
+
+causal-curl-embedding-endpoint: causal-run-original-embeddings
+	@./scripts/utils/curl-embedding-server.sh | ./scripts/causal/compare-embeddings-logits.sh
+
+causal-quantize-Q8_0: QUANTIZED_TYPE = Q8_0
+causal-quantize-Q8_0: causal-quantize-model
+
+causal-quantize-Q4_0: QUANTIZED_TYPE = Q4_0
+causal-quantize-Q4_0: causal-quantize-model
+
+# For Quantization Aware Trained (QAT) models in Q4_0 we explicitly set the
+# token embedding and output types to Q8_0 instead of the default Q6_K.
+causal-quantize-qat-Q4_0: QUANTIZED_TYPE = Q4_0
+causal-quantize-qat-Q4_0: TOKEN_EMBD_TYPE = Q8_0
+causal-quantize-qat-Q4_0: OUTPUT_TYPE = Q8_0
+causal-quantize-qat-Q4_0: causal-quantize-model
+
+causal-quantize-model:
+	$(call quantize_model,$(CONVERTED_MODEL),QUANTIZED_MODEL)
+
+causal-run-quantized-model:
+	@QUANTIZED_MODEL="$(QUANTIZED_MODEL)" ./scripts/causal/run-converted-model.sh ${QUANTIZED_MODEL}
+
+
+###
+### Embedding Model targets/recipes
+###
+
+embedding-convert-model-bf16: OUTTYPE=bf16
+embedding-convert-model-bf16: embedding-convert-model
+
+embedding-convert-model:
+	$(call validate_embedding_model_path,embedding-convert-model)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(EMBEDDING_MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/embedding/convert-model.sh
+
+embedding-run-original-model:
+	$(call validate_embedding_model_path,embedding-run-original-model)
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/embedding/run-original-model.py
+
+embedding-run-converted-model:
+	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/embedding/run-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-verify-logits: embedding-run-original-model embedding-run-converted-model
+	@./scripts/embedding/compare-embeddings-logits.sh
+
+embedding-inspect-original-model:
+	$(call validate_embedding_model_path,embedding-inspect-original-model)
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/utils/inspect-org-model.py -m ${EMBEDDING_MODEL_PATH}
+
+embedding-inspect-converted-model:
+	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/utils/inspect-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-start-embedding-server:
+	@./scripts/utils/run-embedding-server.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-curl-embedding-endpoint:
+	@./scripts/utils/curl-embedding-server.sh | ./scripts/embedding/compare-embeddings-logits.sh
+
+embedding-quantize-Q8_0: QUANTIZED_TYPE = Q8_0
+embedding-quantize-Q8_0: embedding-quantize-model
+
+embedding-quantize-Q4_0: QUANTIZED_TYPE = Q4_0
+embedding-quantize-Q4_0: embedding-quantize-model
+
+# For Quantization Aware Trained (QAT) models in Q4_0 we explicitly set the
+# token embedding and output types to Q8_0 instead of the default Q6_K.
+embedding-quantize-qat-Q4_0: QUANTIZED_TYPE = Q4_0
+embedding-quantize-qat-Q4_0: TOKEN_EMBD_TYPE = Q8_0
+embedding-quantize-qat-Q4_0: OUTPUT_TYPE = Q8_0
+embedding-quantize-qat-Q4_0: embedding-quantize-model
+
+embedding-quantize-model:
+	$(call quantize_model,$(CONVERTED_EMBEDDING_MODEL),QUANTIZED_EMBEDDING_MODEL)
+
+embedding-run-quantized-model:
+	@./scripts/embedding/run-converted-model.sh ${QUANTIZED_EMBEDDING_MODEL}
+
+###
+### Perplexity targets/recipes
+###
+perplexity-data-gen:
+	CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/utils/perplexity-gen.sh
+
+perplexity-run-full:
+	QUANTIZED_MODEL="$(QUANTIZED_MODEL)" LOOGITS_FILE="$(LOGITS_FILE)" \
+	./scripts/utils/perplexity-run.sh
+
+perplexity-run:
+	QUANTIZED_MODEL="$(QUANTIZED_MODEL)" ./scripts/utils/perplexity-run-simple.sh
+
+###
+### HuggingFace targets/recipes
+###
+
+hf-create-model:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}"
+
+hf-create-model-dry-run:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -d
+
+hf-create-model-embedding:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -e
+
+hf-create-model-embedding-dry-run:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -e -d
+
+hf-create-model-private:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -p
+
+hf-upload-gguf-to-model:
+	@./scripts/utils/hf-upload-gguf-model.py -m "${MODEL_PATH}" -r "${REPO_ID}" -o "${NAME_IN_REPO}"
+
+hf-create-collection:
+	@./scripts/utils/hf-create-collection.py -n "${NAME}" -d "${DESCRIPTION}" -ns "${NAMESPACE}"
+
+hf-add-model-to-collection:
+	@./scripts/utils/hf-add-model-to-collection.py -c "${COLLECTION}" -m "${MODEL}"
+
+
+.PHONY: clean
+clean:
+	@${RM} -rf data .converted_embedding_model.txt .converted_model.txt .embedding_model_name.txt .model_name.txt
+

examples/model-conversion/README.md

@@ -0,0 +1,367 @@
+# Model Conversion Example
+This directory contains scripts and code to help in the process of converting
+HuggingFace PyTorch models to GGUF format.
+
+The motivation for having this is that the conversion process can often be an
+iterative process, where the original model is inspected, converted, updates
+made to llama.cpp, converted again, etc. Once the model has been converted it
+needs to be verified against the original model, and then optionally quantified,
+and in some cases perplexity checked of the quantized model. And finally the
+model/models need to the ggml-org on Hugging Face. This tool/example tries to
+help with this process.
+
+### Overview
+The idea is that the makefile targets and scripts here can be used in the
+development/conversion process assisting with things like:
+
+* inspect/run the original model to figure out how it works
+* convert the original model to GGUF format
+* inspect/run the converted model
+* verify the logits produced by the original model and the converted model
+* quantize the model to GGUF format
+* run perplexity evaluation to verify that the quantized model is performing
+  as expected
+* upload the model to HuggingFace to make it available for others
+
+## Setup
+Create virtual python environment
+```console
+$ python3.11 -m venv venv
+$ source venv/bin/activate
+(venv) $ pip install -r requirements.txt
+```
+
+## Causal Language Model Conversion
+This section describes the steps to convert a causal language model to GGUF and
+to verify that the conversion was successful.
+
+### Download the original model
+First, clone the original model to some local directory:
+```console
+$ mkdir models && cd models
+$ git clone https://huggingface.co/user/model_name
+$ cd model_name
+$ git lfs install
+$ git lfs pull
+```
+
+### Set the MODEL_PATH
+The path to the downloaded model can be provided in two ways:
+
+**Option 1: Environment variable (recommended for iterative development)**
+```console
+export MODEL_PATH=~/work/ai/models/some_model
+```
+
+**Option 2: Command line argument (for one-off tasks)**
+```console
+make causal-convert-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+Command line arguments take precedence over environment variables when both are provided.
+
+In cases where the transformer implementation for the model has not been released
+yet it is possible to set the environment variable `UNRELEASED_MODEL_NAME` which
+will then cause the transformer implementation to be loaded explicitely and not
+use AutoModelForCausalLM:
+```
+export UNRELEASED_MODEL_NAME=SomeNewModel
+```
+
+### Inspecting the original tensors
+```console
+# Using environment variable
+(venv) $ make causal-inspect-original-model
+
+# Or using command line argument
+(venv) $ make causal-inspect-original-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+### Running the original model
+This is mainly to verify that the original model works, and to compare the output
+from the converted model.
+```console
+# Using environment variable
+(venv) $ make causal-run-original-model
+
+# Or using command line argument
+(venv) $ make causal-run-original-model MODEL_PATH=~/work/ai/models/some_model
+```
+This command will save two files to the `data` directory, one is a binary file
+containing logits which will be used for comparison with the converted model
+later, and the other is a text file which allows for manual visual inspection.
+
+### Model conversion
+After updates have been made to [gguf-py](../../gguf-py) to add support for the
+new model, the model can be converted to GGUF format using the following command:
+```console
+# Using environment variable
+(venv) $ make causal-convert-model
+
+# Or using command line argument
+(venv) $ make causal-convert-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+### Inspecting the converted model
+The converted model can be inspected using the following command:
+```console
+(venv) $ make inspect-converted-model
+```
+
+### Running the converted model
+```console
+(venv) $ make run-converted-model
+```
+
+### Model logits verfication
+The following target will run the original model and the converted model and
+compare the logits:
+```console
+(venv) $ make causal-verify-logits
+```
+
+### Quantizing the model
+The causal model can be quantized to GGUF format using the following command:
+```console
+(venv) $ make causal-quantize-Q8_0
+Quantized model saved to: /path/to/quantized/model-Q8_0.gguf
+Export the quantized model path to QUANTIZED_MODEL variable in your environment
+```
+This will show the path to the quantized model in the terminal, which can then
+be used to set the `QUANTIZED_MODEL` environment variable:
+```console
+export QUANTIZED_MODEL=/path/to/quantized/model-Q8_0.gguf
+```
+Then the quantized model can be run using the following command:
+```console
+(venv) $ make causal-run-quantized-model
+```
+
+### Quantizing QAT (Quantization Aware Training) models
+When quantizing to `Q4_0`, the default data type for the token embedding weights
+will be `Q6_K`. For models that are going to be uploaded to ggml-org it is
+recommended to use `Q8_0` instead for the embeddings and output tensors.
+The reason is that although `Q6_K` is smaller in size, it requires more compute
+to unpack, which can hurt performance during output generation when the entire
+embedding matrix must be dequantized to compute vocabulary logits. `Q8_0`
+provides practically full quality with better computational efficiency.
+```console
+(venv) $ make causal-quantize-qat-Q4_0
+```
+
+
+## Embedding Language Model Conversion
+
+### Download the original model
+```console
+$ mkdir models && cd models
+$ git clone https://huggingface.co/user/model_name
+$ cd model_name
+$ git lfs install
+$ git lfs pull
+```
+
+The path to the embedding model can be provided in two ways:
+
+**Option 1: Environment variable (recommended for iterative development)**
+```console
+export EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+
+**Option 2: Command line argument (for one-off tasks)**
+```console
+make embedding-convert-model EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+
+Command line arguments take precedence over environment variables when both are provided.
+
+### Running the original model
+This is mainly to verify that the original model works and to compare the output
+with the output from the converted model.
+```console
+# Using environment variable
+(venv) $ make embedding-run-original-model
+
+# Or using command line argument
+(venv) $ make embedding-run-original-model EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+This command will save two files to the `data` directory, one is a binary
+file containing logits which will be used for comparison with the converted
+model, and the other is a text file which allows for manual visual inspection.
+
+### Model conversion
+After updates have been made to [gguf-py](../../gguf-py) to add support for the
+new model the model can be converted to GGUF format using the following command:
+```console
+(venv) $ make embedding-convert-model
+```
+
+### Run the converted model
+```console
+(venv) $ make embedding-run-converted-model
+```
+
+### Model logits verfication
+The following target will run the original model and the converted model (which
+was done manually in the previous steps) and compare the logits:
+```console
+(venv) $ make embedding-verify-logits
+```
+
+### llama-server verification
+To verify that the converted model works with llama-server, the following
+command can be used:
+```console
+(venv) $ make embedding-start-embedding-server
+```
+Then open another terminal and set the `EMBEDDINGS_MODEL_PATH` environment
+variable as this will not be inherited by the new terminal:
+```console
+(venv) $ make embedding-curl-embedding-endpoint
+```
+This will call the `embedding` endpoing and the output will be piped into
+the same verification script as used by the target `embedding-verify-logits`.
+
+The causal model can also be used to produce embeddings and this can be verified
+using the following commands:
+```console
+(venv) $ make causal-start-embedding-server
+```
+Then open another terminal and set the `MODEL_PATH` environment
+variable as this will not be inherited by the new terminal:
+```console
+(venv) $ make casual-curl-embedding-endpoint
+```
+
+### Quantizing the model
+The embedding model can be quantized to GGUF format using the following command:
+```console
+(venv) $ make embedding-quantize-Q8_0
+Quantized model saved to: /path/to/quantized/model-Q8_0.gguf
+Export the quantized model path to QUANTIZED_EMBEDDING_MODEL variable in your environment
+```
+This will show the path to the quantized model in the terminal, which can then
+be used to set the `QUANTIZED_EMBEDDING_MODEL` environment variable:
+```console
+export QUANTIZED_EMBEDDING_MODEL=/path/to/quantized/model-Q8_0.gguf
+```
+Then the quantized model can be run using the following command:
+```console
+(venv) $ make embedding-run-quantized-model
+```
+
+### Quantizing QAT (Quantization Aware Training) models
+When quantizing to `Q4_0`, the default data type for the token embedding weights
+will be `Q6_K`. For models that are going to be uploaded to ggml-org it is
+recommended to use `Q8_0` instead for the embeddings and output tensors.
+The reason is that although `Q6_K` is smaller in size, it requires more compute
+to unpack, which can hurt performance during output generation when the entire
+embedding matrix must be dequantized to compute vocabulary logits. `Q8_0`
+provides practically full quality with better computational efficiency.
+```console
+(venv) $ make embedding-quantize-qat-Q4_0
+```
+
+## Perplexity Evaluation
+
+### Simple perplexity evaluation
+This allows to run the perplexity evaluation without having to generate a
+token/logits file:
+```console
+(venv) $ make perplexity-run QUANTIZED_MODEL=~/path/to/quantized/model.gguf
+```
+This will use the wikitext dataset to run the perplexity evaluation and
+output the perplexity score to the terminal. This value can then be compared
+with the perplexity score of the unquantized model.
+
+### Full perplexity evaluation
+First use the converted, non-quantized, model to generate the perplexity evaluation
+dataset using the following command:
+```console
+$ make perplexity-data-gen CONVERTED_MODEL=~/path/to/converted/model.gguf
+```
+This will generate a file in the `data` directory named after the model and with
+a `.kld` suffix which contains the tokens and the logits for the wikitext dataset.
+
+After the dataset has been generated, the perplexity evaluation can be run using
+the quantized model:
+```console
+$ make perplexity-run-full QUANTIZED_MODEL=~/path/to/quantized/model-Qxx.gguf LOGITS_FILE=data/model.gguf.ppl
+```
+
+> 📝 **Note:** The `LOGITS_FILE` is the file generated by the previous command
+> can be very large, so make sure you have enough disk space available.
+
+## HuggingFace utilities
+The following targets are useful for creating collections and model repositories
+on Hugging Face in the the ggml-org. These can be used when preparing a relase
+to script the process for new model releases.
+
+For the following targets a `HF_TOKEN` environment variable is required.
+
+> 📝 **Note:** Don't forget to logout from Hugging Face after running these
+> commands, otherwise you might have issues pulling/cloning repositories as
+> the token will still be in use:
+> $ huggingface-cli logout
+> $ unset HF_TOKEN
+
+### Create a new Hugging Face Model (model repository)
+This will create a new model repsository on Hugging Face with the specified
+model name.
+```console
+(venv) $ make hf-create-model MODEL_NAME='TestModel' NAMESPACE="danbev" ORIGINAL_BASE_MODEL="some-base-model"
+Repository ID:  danbev/TestModel-GGUF
+Repository created: https://huggingface.co/danbev/TestModel-GGUF
+```
+Note that we append a `-GGUF` suffix to the model name to ensure a consistent
+naming convention for GGUF models.
+
+An embedding model can be created using the following command:
+```console
+(venv) $ make hf-create-model-embedding MODEL_NAME='TestEmbeddingModel' NAMESPACE="danbev" ORIGINAL_BASE_MODEL="some-base-model"
+```
+The only difference is that the model card for an embedding model will be different
+with regards to the llama-server command and also how to access/call the embedding
+endpoint.
+
+### Upload a GGUF model to model repository
+The following target uploads a model to an existing Hugging Face model repository.
+```console
+(venv) $ make hf-upload-gguf-to-model MODEL_PATH=dummy-model1.gguf REPO_ID=danbev/TestModel-GGUF
+📤 Uploading dummy-model1.gguf to danbev/TestModel-GGUF/dummy-model1.gguf
+✅ Upload successful!
+🔗 File available at: https://huggingface.co/danbev/TestModel-GGUF/blob/main/dummy-model1.gguf
+```
+This command can also be used to update an existing model file in a repository.
+
+### Create a new Collection
+```console
+(venv) $ make hf-new-collection NAME=TestCollection DESCRIPTION="Collection for testing scripts" NAMESPACE=danbev
+🚀 Creating Hugging Face Collection
+Title: TestCollection
+Description: Collection for testing scripts
+Namespace: danbev
+Private: False
+✅ Authenticated as: danbev
+📚 Creating collection: 'TestCollection'...
+✅ Collection created successfully!
+📋 Collection slug: danbev/testcollection-68930fcf73eb3fc200b9956d
+🔗 Collection URL: https://huggingface.co/collections/danbev/testcollection-68930fcf73eb3fc200b9956d
+
+🎉 Collection created successfully!
+Use this slug to add models: danbev/testcollection-68930fcf73eb3fc200b9956d
+```
+
+### Add model to a Collection
+```console
+(venv) $ make hf-add-model-to-collection COLLECTION=danbev/testcollection-68930fcf73eb3fc200b9956d MODEL=danbev/TestModel-GGUF
+✅ Authenticated as: danbev
+🔍 Checking if model exists: danbev/TestModel-GGUF
+✅ Model found: danbev/TestModel-GGUF
+📚 Adding model to collection...
+✅ Model added to collection successfully!
+🔗 Collection URL: https://huggingface.co/collections/danbev/testcollection-68930fcf73eb3fc200b9956d
+
+🎉 Model added successfully!
+
+```

examples/model-conversion/logits.cpp

@@ -0,0 +1,210 @@
+#include "llama.h"
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <ctype.h>
+#include <filesystem>
+
+static void print_usage(int, char ** argv) {
+    printf("\nexample usage:\n");
+    printf("\n    %s -m model.gguf [-ngl n_gpu_layers] -embd-mode [prompt]\n", argv[0]);
+    printf("\n");
+}
+
+int main(int argc, char ** argv) {
+    std::string model_path;
+    std::string prompt = "Hello, my name is";
+    int ngl = 0;
+    bool embedding_mode = false;
+
+    {
+        int i = 1;
+        for (; i < argc; i++) {
+            if (strcmp(argv[i], "-m") == 0) {
+                if (i + 1 < argc) {
+                    model_path = argv[++i];
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else if (strcmp(argv[i], "-ngl") == 0) {
+                if (i + 1 < argc) {
+                    try {
+                        ngl = std::stoi(argv[++i]);
+                    } catch (...) {
+                        print_usage(argc, argv);
+                        return 1;
+                    }
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else if (strcmp(argv[i], "-embd-mode") == 0) {
+                if (i + 1 < argc) {
+                    try {
+                        embedding_mode = true;
+                    } catch (...) {
+                        print_usage(argc, argv);
+                        return 1;
+                    }
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else {
+                // prompt starts here
+                break;
+            }
+        }
+
+        if (model_path.empty()) {
+            print_usage(argc, argv);
+            return 1;
+        }
+
+        if (i < argc) {
+            prompt = argv[i++];
+            for (; i < argc; i++) {
+                prompt += " ";
+                prompt += argv[i];
+            }
+        }
+    }
+
+    ggml_backend_load_all();
+    llama_model_params model_params = llama_model_default_params();
+    model_params.n_gpu_layers = ngl;
+
+    llama_model * model = llama_model_load_from_file(model_path.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // Extract basename from model_path
+    const char * basename = strrchr(model_path.c_str(), '/');
+    basename = (basename == NULL) ? model_path.c_str() : basename + 1;
+
+    char model_name[256];
+    strncpy(model_name, basename, 255);
+    model_name[255] = '\0';
+
+    char * dot = strrchr(model_name, '.');
+    if (dot != NULL && strcmp(dot, ".gguf") == 0) {
+        *dot = '\0';
+    }
+    printf("Model name: %s\n", model_name);
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+    const int n_prompt = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, true, true);
+
+    std::vector<llama_token> prompt_tokens(n_prompt);
+    if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true, true) < 0) {
+        fprintf(stderr, "%s: error: failed to tokenize the prompt\n", __func__);
+        return 1;
+    }
+
+    llama_context_params ctx_params = llama_context_default_params();
+    ctx_params.n_ctx = n_prompt;
+    ctx_params.n_batch = n_prompt;
+    ctx_params.no_perf = false;
+    if (embedding_mode) {
+        ctx_params.embeddings = true;
+        ctx_params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+        ctx_params.n_ubatch = ctx_params.n_batch;
+    }
+
+    llama_context * ctx = llama_init_from_model(model, ctx_params);
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    printf("Input prompt: \"%s\"\n", prompt.c_str());
+    printf("Tokenized prompt (%d tokens): ", n_prompt);
+    for (auto id : prompt_tokens) {
+        char buf[128];
+        int n = llama_token_to_piece(vocab, id, buf, sizeof(buf), 0, true);
+        if (n < 0) {
+            fprintf(stderr, "%s: error: failed to convert token to piece\n", __func__);
+            return 1;
+        }
+        std::string s(buf, n);
+        printf("%s", s.c_str());
+    }
+    printf("\n");
+
+    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
+
+    if (llama_decode(ctx, batch)) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return 1;
+    }
+
+    float * logits;
+    int n_logits;
+    const char * type;
+
+    if (embedding_mode) {
+        logits = llama_get_embeddings(ctx);
+        n_logits = llama_model_n_embd(model) * batch.n_tokens;
+        type = "-embeddings";
+        printf("Embeddings size: %d\n", n_logits);
+    } else {
+        logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
+        n_logits = llama_vocab_n_tokens(vocab);
+        type = "";
+        printf("Vocab size: %d\n", n_logits);
+    }
+
+    std::filesystem::create_directory("data");
+
+    // Save logits to binary file
+    char bin_filename[512];
+    snprintf(bin_filename, sizeof(bin_filename), "data/llamacpp-%s%s.bin", model_name, type);
+    printf("Saving logits to %s\n", bin_filename);
+
+    FILE * f = fopen(bin_filename, "wb");
+    if (f == NULL) {
+        fprintf(stderr, "%s: error: failed to open binary output file\n", __func__);
+        return 1;
+    }
+    fwrite(logits, sizeof(float), n_logits, f);
+    fclose(f);
+
+    // Also save as text for debugging
+    char txt_filename[512];
+    snprintf(txt_filename, sizeof(txt_filename), "data/llamacpp-%s%s.txt", model_name, type);
+    f = fopen(txt_filename, "w");
+    if (f == NULL) {
+        fprintf(stderr, "%s: error: failed to open text output file\n", __func__);
+        return 1;
+    }
+    for (int i = 0; i < n_logits; i++) {
+        fprintf(f, "%d: %.6f\n", i, logits[i]);  // Added index and changed format
+    }
+    fclose(f);
+
+    // Print first and last 10 logits for quick verification
+    printf("First 10 logits: ");
+    for (int i = 0; i < 10 && i < n_logits; i++) {
+        printf("%.6f ", logits[i]);
+    }
+    printf("\n");
+
+    printf("Last 10 logits: ");
+    for (int i = n_logits - 10; i < n_logits; i++) {
+        if (i >= 0) printf("%.6f ", logits[i]);
+    }
+    printf("\n\n");
+
+    printf("Logits saved to %s\n", bin_filename);
+    printf("Logits saved to %s\n", txt_filename);
+
+    llama_free(ctx);
+    llama_model_free(model);
+
+    return 0;
+}

examples/model-conversion/requirements.txt

@@ -0,0 +1,5 @@
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch~=2.6.0
+torchvision~=0.21.0
+transformers~=4.55.0
+huggingface-hub~=0.34.0

examples/model-conversion/scripts/causal/compare-embeddings-logits.sh

@@ -0,0 +1,43 @@
+#/bin/bash
+
+set -e
+
+MODEL_PATH="${1:-"$MODEL_PATH"}"
+MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
+
+if [ -t 0 ]; then
+    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
+else
+    # Process piped JSON data and convert to binary (matching logits.cpp format)
+    TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)
+    python3 -c "
+import json
+import sys
+import struct
+
+data = json.load(sys.stdin)
+
+# Flatten all embeddings completely
+flattened = []
+for item in data:
+    embedding = item['embedding']
+    for token_embedding in embedding:
+        flattened.extend(token_embedding)
+
+print(f'Total embedding values: {len(flattened)}', file=sys.stderr)
+
+# Write as binary floats - matches logitc.cpp fwrite format
+with open('$TEMP_FILE', 'wb') as f:
+    for value in flattened:
+        f.write(struct.pack('f', value))
+"
+    CPP_EMBEDDINGS="$TEMP_FILE"
+    trap "rm -f $TEMP_FILE" EXIT
+fi
+
+python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
+    --python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \
+    --cpp-embeddings $CPP_EMBEDDINGS \
+    --prompt "Hello world today" \
+    --causal
+

examples/model-conversion/scripts/causal/compare-logits.py

@@ -0,0 +1,88 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import sys
+import os
+from pathlib import Path
+
+def quick_logits_check(pytorch_file, llamacpp_file):
+    """Lightweight sanity check before NMSE"""
+
+    try:
+        pytorch_logits = np.fromfile(pytorch_file, dtype=np.float32)
+        llamacpp_logits = np.fromfile(llamacpp_file, dtype=np.float32)
+    except Exception as e:
+        print(f"❌ NOK: Failed to load files - {e}")
+        return False
+
+    # Check shapes match
+    if pytorch_logits.shape != llamacpp_logits.shape:
+        print(f"❌ NOK: Shape mismatch - PyTorch: {pytorch_logits.shape}, llama.cpp: {llamacpp_logits.shape}")
+        return False
+
+    # Calculate key metrics
+    diff = pytorch_logits - llamacpp_logits
+    abs_diff = np.abs(diff)
+    max_diff = np.max(abs_diff)
+
+    # Get top 10 predictions from both models
+    pytorch_top10 = np.argsort(pytorch_logits)[-10:][::-1]
+    llamacpp_top10 = np.argsort(llamacpp_logits)[-10:][::-1]
+    print(f"Top 10 PyTorch logits: {pytorch_logits[pytorch_top10]}")
+    print(f"Top 10 llama.cpp logits: {llamacpp_logits[llamacpp_top10]}")
+    print(f"Max absolute difference: {max_diff:.4f}")
+
+    if max_diff > 1.0:
+        print(f"❌ NOK: Large differences detected - max diff: {max_diff:.4f}")
+        return False
+
+    return True
+
+def main():
+    model_path = os.getenv('MODEL_PATH')
+    if not model_path:
+        print("Error: MODEL_PATH environment variable not set")
+        sys.exit(1)
+
+    if not os.path.exists(model_path):
+        print(f"Error: Model file not found: {model_path}")
+        sys.exit(1)
+
+    model_name = os.path.splitext(os.path.basename(model_path))[0]
+    data_dir = Path("data")
+
+    pytorch_file = data_dir / f"pytorch-{model_name}.bin"
+    llamacpp_file = data_dir / f"llamacpp-{model_name}.bin"
+
+    if not pytorch_file.exists():
+        print(f"Error: PyTorch logits file not found: {pytorch_file}")
+        print("Please run scripts/run-org-model.sh first to generate this file.")
+        sys.exit(1)
+
+    if not llamacpp_file.exists():
+        print(f"Error: llama.cpp logits file not found: {llamacpp_file}")
+        print("Please run scripts/run-converted-model.sh first to generate this file.")
+        sys.exit(1)
+
+    print("Checked all required files were found. Proceeding...\n")
+
+
+    print("🔍 GGML Model Validation for model ", model_name)
+    print("=" * 40)
+    print(f"PyTorch logits  : {pytorch_file}")
+    print(f"llama.cpp logits: {llamacpp_file}")
+    print()
+
+    success = quick_logits_check(pytorch_file, llamacpp_file)
+
+    # Exit with appropriate code
+    if success:
+        print("✅ OK: Lightweight model check successful!")
+        print("       Ok to proceed with NMSE check...")
+        sys.exit(0)
+    else:
+        print(f"❌ NOK: Top 10 predictions don't match - generation will differ")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/causal/convert-model.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+
+MODEL_NAME="${MODEL_NAME:-$(basename "$MODEL_PATH")}"
+OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
+TYPE="${OUTTYPE:-f16}"
+METADATA_OVERRIDE="${METADATA_OVERRIDE:-}"
+CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+
+echo "Model path: ${MODEL_PATH}"
+echo "Model name: ${MODEL_NAME}"
+echo "Data  type: ${TYPE}"
+echo "Converted model path:: ${CONVERTED_MODEL}"
+echo "Metadata override: ${METADATA_OVERRIDE}"
+python ../../convert_hf_to_gguf.py --verbose \
+    ${MODEL_PATH} \
+    --outfile ${CONVERTED_MODEL} \
+    --outtype ${TYPE} \
+    --metadata "${METADATA_OVERRIDE}"
+
+echo ""
+echo "The environment variable CONVERTED_MODEL can be set to this path using:"
+echo "export CONVERTED_MODEL=$(realpath ${CONVERTED_MODEL})"

examples/model-conversion/scripts/causal/modelcard.template

@@ -0,0 +1,13 @@
+---
+base_model:
+- {base_model}
+---
+# {model_name} GGUF
+
+Recommended way to run this model:
+
+```sh
+llama-server -hf {namespace}/{model_name}-GGUF -c 0 -fa
+```
+
+Then, access http://localhost:8080

examples/model-conversion/scripts/causal/run-casual-gen-embeddings-org.sh

@@ -0,0 +1,113 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import importlib
+import sys
+import torch
+import numpy as np
+
+from transformers import AutoTokenizer, AutoConfig, AutoModel, AutoModelForCausalLM
+from pathlib import Path
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+config = AutoConfig.from_pretrained(model_path)
+
+print("Model type:       ", config.model_type)
+print("Vocab size:       ", config.vocab_size)
+print("Hidden size:      ", config.hidden_size)
+print("Number of layers: ", config.num_hidden_layers)
+print("BOS token id:     ", config.bos_token_id)
+print("EOS token id:     ", config.eos_token_id)
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}ForCausalLM"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+else:
+    model = AutoModelForCausalLM.from_pretrained(model_path)
+print(f"Model class: {type(model)}")
+#print(f"Model file: {type(model).__module__}")
+
+model_name = os.path.basename(model_path)
+print(f"Model name: {model_name}")
+
+prompt = "Hello world today"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+print(f"Input tokens: {input_ids}")
+print(f"Input text: {repr(prompt)}")
+print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+
+with torch.no_grad():
+    outputs = model(input_ids, output_hidden_states=True)
+
+    # Extract hidden states from the last layer
+    # outputs.hidden_states is a tuple of (num_layers + 1) tensors
+    # Index -1 gets the last layer, shape: [batch_size, seq_len, hidden_size]
+    last_hidden_states = outputs.hidden_states[-1]
+
+    # Get embeddings for all tokens
+    token_embeddings = last_hidden_states[0].cpu().numpy()  # Remove batch dimension
+
+    print(f"Hidden states shape: {last_hidden_states.shape}")
+    print(f"Token embeddings shape: {token_embeddings.shape}")
+    print(f"Hidden dimension: {token_embeddings.shape[-1]}")
+    print(f"Number of tokens: {token_embeddings.shape[0]}")
+
+    # Save raw token embeddings
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
+
+    # Save all token embeddings as binary
+    print(token_embeddings)
+    token_embeddings.astype(np.float32).tofile(bin_filename)
+
+    # Save as text for inspection
+    with open(txt_filename, "w") as f:
+        for i, embedding in enumerate(token_embeddings):
+            for j, val in enumerate(embedding):
+                f.write(f"{i} {j} {val:.6f}\n")
+
+    # Print embeddings per token in the requested format
+    print("\nToken embeddings:")
+    tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
+    for i, embedding in enumerate(token_embeddings):
+        # Format: show first few values, ..., then last few values
+        if len(embedding) > 10:
+            # Show first 3 and last 3 values with ... in between
+            first_vals = " ".join(f"{val:8.6f}" for val in embedding[:3])
+            last_vals = " ".join(f"{val:8.6f}" for val in embedding[-3:])
+            print(f"embedding {i}: {first_vals}  ... {last_vals}")
+        else:
+            # If embedding is short, show all values
+            vals = " ".join(f"{val:8.6f}" for val in embedding)
+            print(f"embedding {i}: {vals}")
+
+    # Also show token info for reference
+    print(f"\nToken reference:")
+    for i, token in enumerate(tokens):
+        print(f"  Token {i}: {repr(token)}")
+
+    print(f"Saved bin logits to: {bin_filename}")
+    print(f"Saved txt logist to: {txt_filename}")

examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m $CONVERTED_MODEL -embd-mode "Hello world today"

examples/model-conversion/scripts/causal/run-converted-model.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" "Hello, my name is"

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

@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import importlib
+from pathlib import Path
+
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+import torch
+import numpy as np
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+config = AutoConfig.from_pretrained(model_path)
+
+print("Model type:       ", config.model_type)
+print("Vocab size:       ", config.vocab_size)
+print("Hidden size:      ", config.hidden_size)
+print("Number of layers: ", config.num_hidden_layers)
+print("BOS token id:     ", config.bos_token_id)
+print("EOS token id:     ", config.eos_token_id)
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+config = AutoConfig.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}ForCausalLM"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)  # Note: from_pretrained, not fromPretrained
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+        exit(1)
+else:
+    model = AutoModelForCausalLM.from_pretrained(model_path)
+
+model_name = os.path.basename(model_path)
+# Printing the Model class to allow for easier debugging. This can be useful
+# when working with models that have not been publicly released yet and this
+# migth require that the concrete class is imported and used directly instead
+# of using AutoModelForCausalLM.
+print(f"Model class: {model.__class__.__name__}")
+
+prompt = "Hello, my name is"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+
+print(f"Input tokens: {input_ids}")
+print(f"Input text: {repr(prompt)}")
+print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+
+with torch.no_grad():
+    outputs = model(input_ids)
+    logits = outputs.logits
+
+    # Extract logits for the last token (next token prediction)
+    last_logits = logits[0, -1, :].cpu().numpy()
+
+    print(f"Logits shape: {logits.shape}")
+    print(f"Last token logits shape: {last_logits.shape}")
+    print(f"Vocab size: {len(last_logits)}")
+
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}.txt"
+
+    # Save to file for comparison
+    last_logits.astype(np.float32).tofile(bin_filename)
+
+    # Also save as text file for easy inspection
+    with open(txt_filename, "w") as f:
+        for i, logit in enumerate(last_logits):
+            f.write(f"{i}: {logit:.6f}\n")
+
+    # Print some sample logits for quick verification
+    print(f"First 10 logits: {last_logits[:10]}")
+    print(f"Last 10 logits: {last_logits[-10:]}")
+
+    # Show top 5 predicted tokens
+    top_indices = np.argsort(last_logits)[-5:][::-1]
+    print("Top 5 predictions:")
+    for idx in top_indices:
+        token = tokenizer.decode([idx])
+        print(f"  Token {idx} ({repr(token)}): {last_logits[idx]:.6f}")
+
+    print(f"Saved bin logits to: {bin_filename}")
+    print(f"Saved txt logist to: {txt_filename}")

examples/model-conversion/scripts/embedding/compare-embeddings-logits.sh

@@ -0,0 +1,42 @@
+#/bin/bash
+
+set -e
+
+MODEL_PATH="${1:-"$EMBEDDING_MODEL_PATH"}"
+MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
+
+if [ -t 0 ]; then
+    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
+else
+    # Process piped JSON data and convert to binary (matching logits.cpp format)
+    TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)
+    python3 -c "
+import json
+import sys
+import struct
+
+data = json.load(sys.stdin)
+
+# Flatten all embeddings completely
+flattened = []
+for item in data:
+    embedding = item['embedding']
+    for token_embedding in embedding:
+        flattened.extend(token_embedding)
+
+print(f'Total embedding values: {len(flattened)}', file=sys.stderr)
+
+# Write as binary floats - matches logitc.cpp fwrite format
+with open('$TEMP_FILE', 'wb') as f:
+    for value in flattened:
+        f.write(struct.pack('f', value))
+"
+    CPP_EMBEDDINGS="$TEMP_FILE"
+    trap "rm -f $TEMP_FILE" EXIT
+fi
+
+python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
+    --python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \
+    --cpp-embeddings $CPP_EMBEDDINGS \
+    --prompt "Hello world today"
+

examples/model-conversion/scripts/embedding/convert-model.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+
+set -e
+
+MODEL_NAME="${MODEL_NAME:-$(basename "$EMBEDDING_MODEL_PATH")}"
+OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
+TYPE="${OUTTYPE:-f16}"
+METADATA_OVERRIDE="${METADATA_OVERRIDE:-}"
+CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+
+echo "Model path: ${EMBEDDING_MODEL_PATH}"
+echo "Model name: ${MODEL_NAME}"
+echo "Data  type: ${TYPE}"
+echo "Converted model path:: ${CONVERTED_MODEL}"
+python ../../convert_hf_to_gguf.py --verbose \
+    ${EMBEDDING_MODEL_PATH} \
+    --outfile ${CONVERTED_MODEL} \
+    --outtype ${TYPE}
+
+echo ""
+echo "The environment variable CONVERTED_EMBEDDING MODEL can be set to this path using:"
+echo "export CONVERTED_EMBEDDING_MODEL=$(realpath ${CONVERTED_MODEL})"

examples/model-conversion/scripts/embedding/modelcard.template

@@ -0,0 +1,48 @@
+---
+base_model:
+- {base_model}
+---
+# {model_name} GGUF
+
+Recommended way to run this model:
+
+```sh
+llama-server -hf {namespace}/{model_name}-GGUF
+```
+
+Then the endpoint can be accessed at http://localhost:8080/embedding, for
+example using `curl`:
+```console
+curl --request POST \
+    --url http://localhost:8080/embedding \
+    --header "Content-Type: application/json" \
+    --data '{{"input": "Hello embeddings"}}' \
+    --silent
+```
+
+Alternatively, the `llama-embedding` command line tool can be used:
+```sh
+llama-embedding -hf {namespace}/{model_name}-GGUF --verbose-prompt -p "Hello embeddings"
+```
+
+#### embd_normalize
+When a model uses pooling, or the pooling method is specified using `--pooling`,
+the normalization can be controlled by the `embd_normalize` parameter.
+
+The default value is `2` which means that the embeddings are normalized using
+the Euclidean norm (L2). Other options are:
+* -1 No normalization
+*  0 Max absolute
+*  1 Taxicab
+*  2 Euclidean/L2
+* \>2 P-Norm
+
+This can be passed in the request body to `llama-server`, for example:
+```sh
+    --data '{{"input": "Hello embeddings", "embd_normalize": -1}}' \
+```
+
+And for `llama-embedding`, by passing `--embd-normalize <value>`, for example:
+```sh
+llama-embedding -hf {namespace}/{model_name}-GGUF  --embd-normalize -1 -p "Hello embeddings"
+```

examples/model-conversion/scripts/embedding/run-converted-model.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_EMBEDDING_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_EMBEDDING_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "Hello world today"

examples/model-conversion/scripts/embedding/run-original-model.py

@@ -0,0 +1,116 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import numpy as np
+import importlib
+from pathlib import Path
+
+from transformers import AutoTokenizer, AutoConfig, AutoModel
+import torch
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('EMBEDDING_MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or EMBEDDING_MODEL_PATH environment variable")
+
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}Model"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)  # Note: from_pretrained, not fromPretrained
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+        exit(1)
+else:
+    model = AutoModel.from_pretrained(model_path)
+print(f"Model class: {type(model)}")
+#print(f"Model file: {type(model).__module__}")
+config = AutoConfig.from_pretrained(model_path)
+
+model_name = os.path.basename(model_path)
+
+texts = [ "Hello world today" ]
+
+encoded = tokenizer(
+    texts,
+    padding=True,
+    truncation=True,
+    return_tensors="pt"
+)
+
+tokens = encoded['input_ids'][0]
+token_strings = tokenizer.convert_ids_to_tokens(tokens)
+for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
+    print(f"{token_id:6d} -> '{token_str}'")
+
+with torch.no_grad():
+    outputs = model(**encoded)
+    hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
+
+    # Extract embeddings for each token (matching LLAMA_POOLING_TYPE_NONE behavior)
+    all_embeddings = hidden_states[0].cpu().numpy()  # Shape: [seq_len, hidden_size]
+
+    print(f"Hidden states shape: {hidden_states.shape}")
+    print(f"All embeddings shape: {all_embeddings.shape}")
+    print(f"Embedding dimension: {all_embeddings.shape[1]}")
+
+    # Print embeddings exactly like embedding.cpp does for LLAMA_POOLING_TYPE_NONE
+    n_embd = all_embeddings.shape[1]
+    n_embd_count = all_embeddings.shape[0]
+
+    print()  # Empty line to match C++ output
+
+    for j in range(n_embd_count):
+        embedding = all_embeddings[j]
+        print(f"embedding {j}: ", end="")
+
+        # Print first 3 values
+        for i in range(min(3, n_embd)):
+            print(f"{embedding[i]:9.6f} ", end="")
+
+        print(" ... ", end="")
+
+        # Print last 3 values
+        for i in range(n_embd - 3, n_embd):
+            print(f"{embedding[i]:9.6f} ", end="")
+
+        print()  # New line
+
+    print()  # Final empty line to match C++ output
+
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
+
+    # Save all embeddings flattened (matching what embedding.cpp would save if it did)
+    flattened_embeddings = all_embeddings.flatten()
+    flattened_embeddings.astype(np.float32).tofile(bin_filename)
+
+    with open(txt_filename, "w") as f:
+        f.write(f"# Model class: {model_name}\n")
+        f.write(f"# Tokens: {token_strings}\n")
+        f.write(f"# Shape: {all_embeddings.shape}\n")
+        f.write(f"# n_embd_count: {n_embd_count}, n_embd: {n_embd}\n\n")
+
+        for j in range(n_embd_count):
+            f.write(f"# Token {j} ({token_strings[j]}):\n")
+            for i, value in enumerate(all_embeddings[j]):
+                f.write(f"{j}_{i}: {value:.6f}\n")
+            f.write("\n")
+    print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} tokens × {n_embd} dimensions)")
+    print("")
+    print(f"Saved bin embeddings to: {bin_filename}")
+    print(f"Saved txt embeddings to: {txt_filename}")

examples/model-conversion/scripts/utils/check-nmse.py

@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import sys
+import os
+import argparse
+from pathlib import Path
+
+def calculate_nmse(reference, test):
+    mse = np.mean((test - reference) ** 2)
+    ref_var = np.var(reference)
+    if ref_var == 0:
+        nmse = float('inf') if mse > 0 else 0.0
+        return mse, mse, ref_var
+
+    nmse = mse / ref_var
+
+    return nmse, mse, ref_var
+
+def load_logits(file_path):
+    if not os.path.exists(file_path):
+        raise FileNotFoundError(f"File not found: {file_path}")
+
+    if file_path.suffix == '.npy':
+        return np.load(file_path)
+    elif file_path.suffix == '.bin':
+        return np.fromfile(file_path, dtype=np.float32)
+    else:
+        # Try to load as text file
+        try:
+            # If it has index format "0: value", extract just values
+            data = []
+            with open(file_path, 'r') as f:
+                for line in f:
+                    if ':' in line:
+                        # Format: "index: value"
+                        value = float(line.split(':')[1].strip())
+                    else:
+                        # Just the value
+                        value = float(line.strip())
+                    data.append(value)
+            return np.array(data, dtype=np.float32)
+        except:
+            return np.loadtxt(file_path, dtype=np.float32)
+
+def interpret_nmse(nmse):
+    """Provide interpretation of NMSE value"""
+    if nmse == 0:
+        return "Perfect match", "🎉"
+    elif nmse < 1e-6:
+        return "Essentially identical", "✅"
+    elif nmse < 1e-4:
+        return "Excellent match", "✅"
+    elif nmse < 1e-3:
+        return "Very good match", "👍"
+    elif nmse < 1e-2:
+        return "Good match", "👍"
+    elif nmse < 0.1:
+        return "Acceptable match", "⚠️"
+    elif nmse < 1.0:
+        return "Poor match", "❌"
+    else:
+        return "Very poor match (worse than noise)", "❌"
+
+def main():
+    parser = argparse.ArgumentParser(description='Validate model logits')
+    parser.add_argument('-m', '--model-path', required=True,  help='Path to the model directory')
+    args = parser.parse_args()
+
+    model_name = os.path.splitext(os.path.basename(args.model_path))[0]
+    data_dir = Path("data")
+
+    pytorch_file = data_dir / f"pytorch-{model_name}.bin"
+    llamacpp_file = data_dir / f"llamacpp-{model_name}.bin"
+
+    print(f"Model name: {model_name}")
+    print(f"PyTorch logits file: {pytorch_file}")
+    print(f"llama.cpp logits file: {llamacpp_file}")
+
+    reference_file = pytorch_file
+    test_file = llamacpp_file
+
+    print("📊 NMSE Check for Model Comparison")
+    print("=" * 50)
+    print(f"Reference (ground truth): {reference_file}")
+    print(f"Test (to evaluate):       {test_file}")
+    print()
+
+    try:
+        print("Loading reference logits...")
+        reference = load_logits(reference_file)
+        print(f"  Shape: {reference.shape}, Type: {reference.dtype}")
+
+        print("Loading test logits...")
+        test = load_logits(test_file)
+        print(f"  Shape: {test.shape}, Type: {test.dtype}")
+
+        # Check shapes match
+        if reference.shape != test.shape:
+            print(f"\n❌ Error: Shape mismatch!")
+            print(f"  Reference: {reference.shape}")
+            print(f"  Test: {test.shape}")
+            sys.exit(1)
+
+        print(f"\n✅ Shapes match: {reference.shape}")
+
+        nmse, mse, ref_var = calculate_nmse(reference, test)
+
+        # Additional metrics
+        max_abs_error = np.max(np.abs(test - reference))
+        mean_abs_error = np.mean(np.abs(test - reference))
+
+        # Results
+        print(f"\n📈 METRICS")
+        print("=" * 30)
+        print(f"MSE (Mean Squared Error):     {mse:.6e}")
+        print(f"Reference Variance:           {ref_var:.6e}")
+        print(f"NMSE:                         {nmse:.6e}")
+        print(f"Max Absolute Error:           {max_abs_error:.6f}")
+        print(f"Mean Absolute Error:          {mean_abs_error:.6f}")
+
+        # NMSE in dB (common in signal processing)
+        if nmse > 0:
+            nmse_db = 10 * np.log10(nmse)
+            print(f"NMSE (dB):                    {nmse_db:.2f} dB")
+
+        # Interpretation
+        interpretation, emoji = interpret_nmse(nmse)
+        print(f"\n🎯 INTERPRETATION")
+        print("=" * 30)
+        print(f"{emoji} {interpretation}")
+
+        # Detailed guidance
+        print(f"\n📋 GUIDANCE")
+        print("=" * 30)
+        if nmse < 1e-3:
+            print("✅ EXCELLENT: Your GGML conversion is working very well!")
+            print("   The differences are negligible for practical use.")
+        elif nmse < 1e-2:
+            print("👍 GOOD: Your GGML conversion is working well.")
+            print("   Small differences are likely due to precision/quantization.")
+        elif nmse < 0.1:
+            print("⚠️  ACCEPTABLE: Conversion is working but with some differences.")
+            print("   Check if you're using quantization (Q4, Q8, etc.)")
+            print("   Test generation quality to see if it's acceptable.")
+        else:
+            print("❌ PROBLEMATIC: Large differences detected.")
+            print("   Check your conversion process for potential issues.")
+            print("   Verify you're using the same model weights.")
+
+        # NMSE benchmarks
+        print(f"\n📚 NMSE BENCHMARKS")
+        print("=" * 30)
+        print("< 1e-6:  Essentially identical")
+        print("< 1e-4:  Excellent (typical for good conversions)")
+        print("< 1e-3:  Very good")
+        print("< 1e-2:  Good (acceptable for most use cases)")
+        print("< 0.1:   Acceptable (may need verification)")
+        print("> 1.0:   Poor (worse than random)")
+
+        # Exit code based on NMSE
+        if nmse < 1e-2:
+            print(f"\n✅ RESULT: PASS (NMSE = {nmse:.2e})")
+            sys.exit(0)
+        else:
+            print(f"\n❌ RESULT: NEEDS REVIEW (NMSE = {nmse:.2e})")
+            sys.exit(1)
+
+    except Exception as e:
+        print(f"❌ Error: {e}")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/utils/create-collection-add-model.sh

@@ -0,0 +1,6 @@
+
+COLLECTION_SLUG=$(python ./create_collection.py --return-slug)
+echo "Created collection: $COLLECTION_SLUG"
+
+# Use it in the next command
+python add_model_to_collection.py "$COLLECTION_SLUG" "username/my-model"

examples/model-conversion/scripts/utils/hf-add-model-to-collection.py

@@ -0,0 +1,80 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import sys
+
+def add_model_to_collection(collection_slug, model_id, note=""):
+    """
+    Add a model to an existing collection
+
+    Args:
+        collection_slug: The slug of the collection (e.g., "username/collection-name-12345")
+        model_id: The model repository ID (e.g., "username/model-name")
+        note: Optional note about the model
+
+    Returns:
+        True if successful, False if failed
+    """
+
+    # Initialize API
+    api = HfApi()
+
+    try:
+        user_info = api.whoami()
+        print(f"✅ Authenticated as: {user_info['name']}")
+
+        # Verify the model exists
+        print(f"🔍 Checking if model exists: {model_id}")
+        try:
+            model_info = api.model_info(model_id)
+        except Exception as e:
+            print(f"❌ Model not found or not accessible: {model_id}")
+            print(f"Error: {e}")
+            return False
+
+        print(f"📚 Adding model to collection...")
+        api.add_collection_item(
+            collection_slug=collection_slug,
+            item_id=model_id,
+            item_type="model",
+            note=note
+        )
+
+        print(f"✅ Model added to collection successfully!")
+        print(f"🔗 Collection URL: https://huggingface.co/collections/{collection_slug}")
+
+        return True
+
+    except Exception as e:
+        print(f"❌ Error adding model to collection: {e}")
+        return False
+
+def main():
+    # This script requires that the environment variable HF_TOKEN is set with your
+    # Hugging Face API token.
+    api = HfApi()
+
+    parser = argparse.ArgumentParser(description='Add model to a Huggingface Collection')
+    parser.add_argument('--collection', '-c', help='The collection slug username/collection-hash', required=True)
+    parser.add_argument('--model', '-m', help='The model to add to the Collection', required=True)
+    parser.add_argument('--note', '-n', help='An optional note/description', required=False)
+    args = parser.parse_args()
+
+    collection = args.collection
+    model = args.model
+    note = args.note
+
+    success = add_model_to_collection(
+        collection_slug=collection,
+        model_id=model,
+        note=note
+    )
+
+    if success:
+        print("\n🎉 Model added successfully!")
+    else:
+        print("\n❌ Failed to add model to collection")
+        sys.exit(1)
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/utils/hf-create-collection.py

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import os
+import sys
+
+
+def create_collection(title, description, private=False, namespace=None, return_slug=False):
+    """
+    Create a new collection on Hugging Face
+
+    Args:
+        title: Collection title
+        description: Collection description
+        private: Whether the collection should be private (default: False)
+        namespace: Optional namespace (defaults to your username)
+
+    Returns:
+        Collection object if successful, None if failed
+    """
+
+    # Check if HF_TOKEN is available
+    token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
+    if not token:
+        print("❌ No HF_TOKEN or HUGGINGFACE_HUB_TOKEN found in environment variables")
+        print("Please set your Hugging Face token as an environment variable")
+        return None
+
+    # Initialize API
+    api = HfApi()
+
+    try:
+        # Test authentication first
+        user_info = api.whoami()
+        if not return_slug:
+            print(f"✅ Authenticated as: {user_info['name']}")
+
+        # Create the collection
+        if not return_slug:
+            print(f"📚 Creating collection: '{title}'...")
+        collection = api.create_collection(
+            title=title,
+            description=description,
+            private=private,
+            namespace=namespace
+        )
+
+        if not return_slug:
+            print(f"✅ Collection created successfully!")
+            print(f"📋 Collection slug: {collection.slug}")
+            print(f"🔗 Collection URL: https://huggingface.co/collections/{collection.slug}")
+
+        return collection
+
+    except Exception as e:
+        print(f"❌ Error creating collection: {e}")
+        return None
+
+def main():
+    # This script requires that the environment variable HF_TOKEN is set with your
+    # Hugging Face API token.
+    api = HfApi()
+
+    parser = argparse.ArgumentParser(description='Create a Huggingface Collection')
+    parser.add_argument('--name', '-n', help='The name/title of the Collection', required=True)
+    parser.add_argument('--description', '-d', help='The description for the Collection', required=True)
+    parser.add_argument('--namespace', '-ns', help='The namespace to add the Collection to', required=True)
+    parser.add_argument('--private', '-p', help='Create a private Collection', action='store_true')  # Fixed
+    parser.add_argument('--return-slug', '-s', help='Only output the collection slug', action='store_true')  # Fixed
+
+    args = parser.parse_args()
+
+    name = args.name
+    description = args.description
+    private = args.private
+    namespace = args.namespace
+    return_slug = args.return_slug
+
+    if not return_slug:
+        print("🚀 Creating Hugging Face Collection")
+        print(f"Title: {name}")
+        print(f"Description: {description}")
+        print(f"Namespace: {namespace}")
+        print(f"Private: {private}")
+
+    collection = create_collection(
+        title=name,
+        description=description,
+        private=private,
+        namespace=namespace,
+        return_slug=return_slug
+    )
+
+    if collection:
+        if return_slug:
+            print(collection.slug)
+        else:
+            print("\n🎉 Collection created successfully!")
+            print(f"Use this slug to add models: {collection.slug}")
+    else:
+        print("\n❌ Failed to create collection")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/utils/hf-create-model.py

@@ -0,0 +1,78 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+
+# This script requires that the environment variable HF_TOKEN is set with your
+# Hugging Face API token.
+api = HfApi()
+
+def load_template_and_substitute(template_path, **kwargs):
+    try:
+        with open(template_path, 'r', encoding='utf-8') as f:
+            template_content = f.read()
+
+        return template_content.format(**kwargs)
+    except FileNotFoundError:
+        print(f"Template file '{template_path}' not found!")
+        return None
+    except KeyError as e:
+        print(f"Missing template variable: {e}")
+        return None
+
+parser = argparse.ArgumentParser(description='Create a new Hugging Face model repository')
+parser.add_argument('--model-name', '-m', help='Name for the model', required=True)
+parser.add_argument('--namespace', '-ns', help='Namespace to add the model to', required=True)
+parser.add_argument('--org-base-model', '-b', help='Original Base model name', default="")
+parser.add_argument('--no-card', action='store_true', help='Skip creating model card')
+parser.add_argument('--private', '-p', action='store_true', help='Create private model')
+parser.add_argument('--embedding', '-e', action='store_true', help='Use embedding model card template')
+parser.add_argument('--dry-run', '-d', action='store_true', help='Print repository info and template without creating repository')
+
+args = parser.parse_args()
+
+repo_id = f"{args.namespace}/{args.model_name}-GGUF"
+print("Repository ID: ", repo_id)
+
+repo_url = None
+if not args.dry_run:
+    repo_url = api.create_repo(
+        repo_id=repo_id,
+        repo_type="model",
+        private=args.private,
+        exist_ok=False
+    )
+
+if not args.no_card:
+    if args.embedding:
+        template_path = "scripts/embedding/modelcard.template"
+    else:
+        template_path = "scripts/causal/modelcard.template"
+
+    print("Template path: ", template_path)
+
+    model_card_content = load_template_and_substitute(
+        template_path,
+        model_name=args.model_name,
+        namespace=args.namespace,
+        base_model=args.org_base_model,
+    )
+
+    if args.dry_run:
+        print("\nTemplate Content:\n")
+        print(model_card_content)
+    else:
+        if model_card_content:
+            api.upload_file(
+                path_or_fileobj=model_card_content.encode('utf-8'),
+                path_in_repo="README.md",
+                repo_id=repo_id
+            )
+            print("Model card created successfully.")
+        else:
+            print("Failed to create model card.")
+
+if not args.dry_run and repo_url:
+    print(f"Repository created: {repo_url}")
+
+

examples/model-conversion/scripts/utils/hf-upload-gguf-model.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import os
+
+def upload_gguf_file(local_file_path, repo_id, filename_in_repo=None):
+    """
+    Upload a GGUF file to a Hugging Face model repository
+
+    Args:
+        local_file_path: Path to your local GGUF file
+        repo_id: Your repository ID (e.g., "username/model-name")
+        filename_in_repo: Optional custom name for the file in the repo
+    """
+
+    if not os.path.exists(local_file_path):
+        print(f"❌ File not found: {local_file_path}")
+        return False
+
+    if filename_in_repo is None:
+        filename_in_repo = os.path.basename(local_file_path)
+
+    if filename_in_repo is None or filename_in_repo == "":
+        filename_in_repo = os.path.basename(local_file_path)
+
+    print(f"📤 Uploading {local_file_path} to {repo_id}/{filename_in_repo}")
+
+    api = HfApi()
+
+    try:
+        api.upload_file(
+            path_or_fileobj=local_file_path,
+            path_in_repo=filename_in_repo,
+            repo_id=repo_id,
+            repo_type="model",
+            commit_message=f"Upload {filename_in_repo}"
+        )
+
+        print("✅ Upload successful!")
+        print(f"🔗 File available at: https://huggingface.co/{repo_id}/blob/main/{filename_in_repo}")
+        return True
+
+    except Exception as e:
+        print(f"❌ Upload failed: {e}")
+        return False
+
+# This script requires that the environment variable HF_TOKEN is set with your
+# Hugging Face API token.
+api = HfApi()
+
+parser = argparse.ArgumentParser(description='Upload a GGUF model to a Huggingface model repository')
+parser.add_argument('--gguf-model-path', '-m', help='The GGUF model file to upload', required=True)
+parser.add_argument('--repo-id', '-r', help='The repository to upload to', required=True)
+parser.add_argument('--name', '-o', help='The name in the model repository', required=False)
+args = parser.parse_args()
+
+upload_gguf_file(args.gguf_model_path, args.repo_id, args.name)

examples/model-conversion/scripts/utils/inspect-converted-model.sh

@@ -0,0 +1,14 @@
+#!/bin/bash
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+../../gguf-py/gguf/scripts/gguf_dump.py $CONVERTED_MODEL

examples/model-conversion/scripts/utils/inspect-org-model.py

@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import json
+from safetensors import safe_open
+from collections import defaultdict
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+# Check if there's an index file (multi-file model)
+index_path = os.path.join(model_path, "model.safetensors.index.json")
+single_file_path = os.path.join(model_path, "model.safetensors")
+
+if os.path.exists(index_path):
+    # Multi-file model
+    print("Multi-file model detected")
+
+    with open(index_path, 'r') as f:
+        index_data = json.load(f)
+
+    # Get the weight map (tensor_name -> file_name)
+    weight_map = index_data.get("weight_map", {})
+
+    # Group tensors by file for efficient processing
+    file_tensors = defaultdict(list)
+    for tensor_name, file_name in weight_map.items():
+        file_tensors[file_name].append(tensor_name)
+
+    print("Tensors in model:")
+
+    # Process each shard file
+    for file_name, tensor_names in file_tensors.items():
+        file_path = os.path.join(model_path, file_name)
+        print(f"\n--- From {file_name} ---")
+
+        with safe_open(file_path, framework="pt") as f:
+            for tensor_name in sorted(tensor_names):
+                tensor = f.get_tensor(tensor_name)
+                print(f"- {tensor_name} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+
+elif os.path.exists(single_file_path):
+    # Single file model (original behavior)
+    print("Single-file model detected")
+
+    with safe_open(single_file_path, framework="pt") as f:
+        keys = f.keys()
+        print("Tensors in model:")
+        for key in sorted(keys):
+            tensor = f.get_tensor(key)
+            print(f"- {key} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+
+else:
+    print(f"Error: Neither 'model.safetensors.index.json' nor 'model.safetensors' found in {model_path}")
+    print("Available files:")
+    if os.path.exists(model_path):
+        for item in sorted(os.listdir(model_path)):
+            print(f"  {item}")
+    else:
+        print(f"  Directory {model_path} does not exist")
+    exit(1)

examples/model-conversion/scripts/utils/perplexity-gen.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+
+set -e
+
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+# Check if data/wikitext-2-raw directory exists
+if [ ! -d "ppl/wikitext-2-raw" ]; then
+    echo "ppl/wikitext-2-raw directory does not exist. Downloading..." >&2
+    mkdir -p ppl
+    pushd ppl
+    ./../../../scripts/get-wikitext-2.sh
+    popd
+fi
+
+mkdir -p ppl
+OUTPUTFILE="ppl/$(basename $CONVERTED_MODEL).kld"
+echo "Model: $CONVERTED_MODEL"
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $CONVERTED_MODEL \
+    -f ppl/wikitext-2-raw/wiki.test.raw \
+    --kl-divergence-base $OUTPUTFILE
+
+echo "Generated logits in $OUTPUTFILE"
+

examples/model-conversion/scripts/utils/perplexity-run-simple.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+
+set -e
+
+QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
+
+if [ -z "$QUANTIZED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. QUANTIZED_MODEL environment variable" >&2
+    exit 1
+fi
+
+# Check if data/wikitext-2-raw directory exists
+if [ ! -d "ppl/wikitext-2-raw" ]; then
+    echo "ppl/wikitext-2-raw directory does not exist. Downloading..." >&2
+    mkdir -p ppl
+    pushd ppl
+    ./../../../scripts/get-wikitext-2.sh
+    popd
+fi
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL -f ppl/wikitext-2-raw/wiki.test.raw
+
+

examples/model-conversion/scripts/utils/perplexity-run.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+
+set -e
+
+QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
+LOGITS_FILE="${1:-"$LOGITS_FILE"}"
+
+if [ -z "$QUANTIZED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. QUANTIZED_MODEL environment variable" >&2
+    exit 1
+fi
+
+if [ ! -f ${LOGITS_FILE} ]; then
+    echo "Error: logits file '${LOGITS_FILE} was not found"
+    echo "Did you run the perplexity-gen.sh script?"
+    exit 1
+fi
+
+echo "Model: $QUANTIZED_MODEL"
+echo "Data file: $LOGITS_FILE"
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL \
+    --kl-divergence-base $LOGITS_FILE \
+    --kl-divergence

examples/model-conversion/scripts/utils/quantize.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+
+set -e
+
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+QUANTIZED_TYPE="${2:-"$QUANTIZED_TYPE"}"
+TOKEN_EMBD_TYPE="${3:-"${TOKEN_EMBD_TYPE}"}"
+OUTPUT_TYPE="${4:-"${OUTPUT_TYPE}"}"
+QUANTIZED_MODEL=$CONVERTED_MODEL
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+if [ -z "$QUANTIZED_TYPE" ]; then
+    echo "Error: QUANTIZED_TYPE is required" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+# Process the quantized model filename
+if [[ "$QUANTIZED_MODEL" == *.gguf ]]; then
+    # Remove .gguf suffix, add quantized type, then add .gguf back
+    BASE_NAME="${QUANTIZED_MODEL%.gguf}"
+    QUANTIZED_MODEL="${BASE_NAME}-${QUANTIZED_TYPE}.gguf"
+else
+    echo "Error: QUANTIZED_MODEL must end with .gguf extension" >&2
+    exit 1
+fi
+
+cmake --build ../../build --target llama-quantize -j8
+
+echo $TOKEN_EMBD_TYPE
+echo $OUTPUT_TYPE
+
+CMD_ARGS=("../../build/bin/llama-quantize")
+[[ -n "$TOKEN_EMBD_TYPE" ]] && CMD_ARGS+=("--token-embedding-type" "$TOKEN_EMBD_TYPE")
+[[ -n "$OUTPUT_TYPE" ]]     && CMD_ARGS+=("--output-tensor-type" "$OUTPUT_TYPE")
+CMD_ARGS+=("$CONVERTED_MODEL" "$QUANTIZED_MODEL" "$QUANTIZED_TYPE")
+
+"${CMD_ARGS[@]}"
+
+echo "Quantized model saved to: $QUANTIZED_MODEL"

examples/model-conversion/scripts/utils/run-embedding-server.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+
+set -e
+#
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-server
+
+../../build/bin/llama-server -m $CONVERTED_MODEL \
+    --embedding \
+    --pooling none

examples/model-conversion/scripts/utils/semantic_check.py

@@ -0,0 +1,179 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import argparse
+import os
+import importlib
+
+from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM, AutoModel
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+def cosine_similarity(a, b=None):
+    a = np.asarray(a)
+    if b is None:
+        b = a
+    else:
+        b = np.asarray(b)
+
+    if a.ndim == 1:
+        a = a.reshape(1, -1)
+    if b.ndim == 1:
+        b = b.reshape(1, -1)
+
+    a_norms = np.linalg.norm(a, axis=1, keepdims=True)
+    b_norms = np.linalg.norm(b, axis=1, keepdims=True)
+
+    a_norms = np.where(a_norms == 0, 1e-8, a_norms)
+    b_norms = np.where(b_norms == 0, 1e-8, b_norms)
+
+    a_normalized = a / a_norms
+    b_normalized = b / b_norms
+
+    # Compute cosine similarity
+    return np.dot(a_normalized, b_normalized.T)
+
+def load_embeddings_from_file(filename, n_tokens, n_embd):
+    embeddings = np.fromfile(filename, dtype=np.float32)
+    return embeddings.reshape(n_tokens, n_embd)
+
+def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
+    np.set_printoptions(suppress=True, precision=6)
+    print("pytorch embeddings:");
+    print(python_emb)
+    print("llama.cpp embeddings:");
+    print(cpp_emb)
+    print(f"\n=== Prompt: '{prompt}' ===")
+    print(f"Tokens: {tokens}")
+    print(f"Embeddings shape: Python {python_emb.shape}, llama.cpp {cpp_emb.shape}")
+
+    n_tokens = len(tokens)
+
+    # 1. Direct embedding comparison
+    print(f"\n1. Raw Embedding Magnitude Comparison:")
+    # Check if the distance of each token embedding from the origin and compare
+    # if the vectors are on the same "sphere". This does not tell us about
+    # direction (meaning of the token embedding), just magnitude.
+    for i in range(n_tokens):
+        py_mag = np.linalg.norm(python_emb[i]) # calculate standard euclidean norm for Python embeddings
+        cpp_mag = np.linalg.norm(cpp_emb[i])   # calculate standard euclidean norm for llama.cpp embeddings
+        ratio = py_mag / cpp_mag if cpp_mag > 0 else float('inf')
+        print(f"   Token {i} ({tokens[i]}): Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
+
+    # 2. Cosine similarity between tokens within each model
+    # Here we check the direction of token embeddings to see if the have the
+    # same meaning (similarity). This is done by calculating cosine similarity
+    # of a pair of token embeddings within each model.
+    print(f"\n2. Within-Model Token Similarities:")
+    print("   Python model:")
+    for i in range(n_tokens):
+        for j in range(i+1, n_tokens):
+            sim = cosine_similarity([python_emb[i]], [python_emb[j]])[0][0]
+            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+
+    print("   llama.cpp model:")
+    for i in range(n_tokens):
+        for j in range(i+1, n_tokens):
+            sim = cosine_similarity([cpp_emb[i]], [cpp_emb[j]])[0][0]
+            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+
+    # 3. Cross-model similarity (same token position)
+    print(f"\n3. Cross-Model Same-Token Similarities:")
+    for i in range(n_tokens):
+        sim = cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0]
+        print(f"   Token {i} ({tokens[i]}): {sim:.4f}")
+
+    # 4. Similarity matrix comparison
+    print(f"\n4. Similarity Matrix Differences:")
+    py_sim_matrix = cosine_similarity(python_emb)
+    cpp_sim_matrix = cosine_similarity(cpp_emb)
+    diff_matrix = np.abs(py_sim_matrix - cpp_sim_matrix)
+
+    print(f"   Max difference: {np.max(diff_matrix):.4f}")
+    print(f"   Mean difference: {np.mean(diff_matrix):.4f}")
+    print(f"   RMS difference: {np.sqrt(np.mean(diff_matrix**2)):.4f}")
+
+    return {
+        'cross_model_similarities': [cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0] for i in range(n_tokens)],
+        'similarity_matrix_diff': diff_matrix,
+        'max_diff': np.max(diff_matrix),
+        'mean_diff': np.mean(diff_matrix),
+        'rms_diff': np.sqrt(np.mean(diff_matrix**2))
+    }
+
+def main():
+    parser = argparse.ArgumentParser(description='Test semantic similarity between Python and llama.cpp embeddings')
+    parser.add_argument('--model-path', '-m', required=True, help='Path to the original Python model')
+    parser.add_argument('--python-embeddings', '-pe', help='Path to pytorch embeddings "logits" binary file')
+    parser.add_argument('--cpp-embeddings', '-ce', help='Path to llama.cpp embeddings "logits" binary file')
+    parser.add_argument('--causal', '-c', default=False, help='if the model is causal (default: false)', action='store_true')
+    parser.add_argument('--prompt', '-p', default='Hello world today', help='Test prompt')
+
+    args = parser.parse_args()
+
+    print("Semantic Similarity Test Between Python and llama.cpp Embedding Models")
+    print("=" * 70)
+
+    # Single prompt detailed comparison
+    print(f"\nTesting with prompt: '{args.prompt}'")
+
+    # Load the python model to get configuration information and also to load the tokenizer.
+    print("Loading model and tokenizer using AutoTokenizer:", args.model_path)
+    tokenizer = AutoTokenizer.from_pretrained(args.model_path)
+    config = AutoConfig.from_pretrained(args.model_path)
+
+    if unreleased_model_name:
+        model_name_lower = unreleased_model_name.lower()
+        unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+        if args.causal:
+            class_name = f"{unreleased_model_name}ForCausalLM"
+        else:
+            class_name = f"{unreleased_model_name}Model"
+        print(f"Model class: {class_name}")
+        print(f"Importing unreleased model module: {unreleased_module_path}")
+
+        try:
+            model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+            model = model_class.from_pretrained(args.model_path)
+        except (ImportError, AttributeError) as e:
+            print(f"Failed to import or load model: {e}")
+            exit(1)
+    else:
+        if args.causal:
+            model = AutoModelForCausalLM.from_pretrained(args.model_path)
+        else:
+            model = AutoModel.from_pretrained(args.model_path)
+
+    encoded = tokenizer(args.prompt, return_tensors="pt")
+    tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])
+    n_tokens = len(tokens)
+    print(f"n_tokens: {n_tokens}");
+    print(f"hidden_size: {model.config.hidden_size}")
+
+    # Load binary embeddings from data directory.
+    llamacpp_embeddings = load_embeddings_from_file(args.cpp_embeddings, n_tokens, model.config.hidden_size)
+    python_embeddings = load_embeddings_from_file(args.python_embeddings, n_tokens, model.config.hidden_size)
+
+    # Run comparison
+    results = test_single_prompt_similarity(python_embeddings, llamacpp_embeddings, tokens, args.prompt)
+
+    # Summary
+    print(f"\n=== SUMMARY ===")
+    avg_cross_sim = np.mean(results['cross_model_similarities'])
+    print(f"Average cross-model similarity: {avg_cross_sim:.4f}")
+    print(f"Similarity matrix RMS difference: {results['rms_diff']:.4f}")
+
+    # Quality assessment
+    if avg_cross_sim > 0.95:
+        print("✅ EXCELLENT: Models are highly similar")
+    elif avg_cross_sim > 0.90:
+        print("✅ VERY GOOD: Models are very similar")
+    elif avg_cross_sim > 0.80:
+        print("⚠️  GOOD: Models are reasonably similar")
+    elif avg_cross_sim > 0.70:
+        print("⚠️  FAIR: Models have some differences")
+    else:
+        print("❌ POOR: Models are significantly different")
+
+if __name__ == "__main__":
+    main()

ggml/include/ggml.h

@@ -512,6 +512,7 @@ extern "C" {
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
         GGML_OP_CONV_2D,
+        GGML_OP_CONV_3D,
         GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
@@ -1940,6 +1941,23 @@ extern "C" {
             int                   d0,  // dilation dimension 0
             int                   d1); // dilation dimension 1
 
+    GGML_API struct ggml_tensor * ggml_conv_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // kernel [KW, KH, KD, IC * OC]
+            struct ggml_tensor  * b,   // input  [W, H, D, C * N]
+            int                   s0,  // stride
+            int                   s1,
+            int                   s2,
+            int                   p0,  // padding
+            int                   p1,
+            int                   p2,
+            int                   d0,  // dilation
+            int                   d1,
+            int                   d2,
+            int                   n_channels,
+            int                   n_batch,
+            int                   n_channels_out);
+
     enum ggml_op_pool {
         GGML_OP_POOL_MAX,
         GGML_OP_POOL_AVG,

ggml/src/ggml-backend.cpp

@@ -1355,15 +1355,15 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
     std::vector<int32_t> ids;
     std::vector<ggml_bitset_t> used_ids;
 
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &splits[i];
+    for (int split_id = 0; split_id < sched->n_splits; split_id++) {
+        struct ggml_backend_sched_split * split = &splits[split_id];
         int split_backend_id = split->backend_id;
         ggml_backend_t split_backend = sched->backends[split_backend_id];
 
         // copy the input tensors to the split backend
-        for (int j = 0; j < split->n_inputs; j++) {
-            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
-            struct ggml_tensor * input = split->inputs[j];
+        for (int input_id = 0; input_id < split->n_inputs; input_id++) {
+            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
+            struct ggml_tensor * input = split->inputs[input_id];
             struct ggml_tensor * input_cpy = tensor_copy(input, split_backend_id, sched->cur_copy);
 
             if (input->flags & GGML_TENSOR_FLAG_INPUT) {
@@ -1398,10 +1398,22 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
 
                     // get the ids
                     ggml_tensor * ids_tensor = node->src[2];
+                    ggml_backend_t ids_backend = split_backend;
+
+                    // if the ids tensor is also an input of the split, it may not have been copied yet to the split backend
+                    // in that case, we use the original ids tensor
+                    for (int i = input_id + 1; i < split->n_inputs; i++) {
+                        if (ids_tensor == tensor_copy(split->inputs[i], split_backend_id, sched->cur_copy)) {
+                            ids_tensor = split->inputs[i];
+                            ids_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[i]);
+                            break;
+                        }
+                    }
+
                     if (ids_tensor != prev_ids_tensor) {
                         ids.resize(ggml_nbytes(ids_tensor) / sizeof(int32_t));
-                        ggml_backend_tensor_get_async(split_backend, ids_tensor, ids.data(), 0, ggml_nbytes(ids_tensor));
-                        ggml_backend_synchronize(split_backend);
+                        ggml_backend_tensor_get_async(ids_backend, ids_tensor, ids.data(), 0, ggml_nbytes(ids_tensor));
+                        ggml_backend_synchronize(ids_backend);
 
                         // find the used experts
                         used_ids.clear();
@@ -1409,6 +1421,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                         for (int64_t i1 = 0; i1 < ids_tensor->ne[1]; i1++) {
                             for (int64_t i0 = 0; i0 < ids_tensor->ne[0]; i0++) {
                                 int32_t id = ids[i1 * ids_tensor->nb[1]/sizeof(int32_t) + i0 * ids_tensor->nb[0]/sizeof(int32_t)];
+                                GGML_ASSERT(id >= 0 && id < n_expert);
                                 ggml_bitset_set(used_ids.data(), id);
                             }
                         }

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -867,6 +867,86 @@ static aclTensor* aclnn_values(ggml_backend_cann_context& ctx, void* buffer,
     return acl_tensor;
 }
 
+/**
+ * @brief Fills a tensor with a scalar value.
+ *
+ * This function fills the destination tensor `acl_dst` with the scalar value
+ * `scalar`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param scalar The scalar value used to fill the tensor.
+ * @param acl_dst The destination tensor to be filled with the scalar value.
+ */
+static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar,
+                              aclTensor* acl_dst) {
+    auto acl_scalar = aclCreateScalar(&scalar, aclDataType::ACL_FLOAT);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceFillScalar, acl_dst, acl_scalar);
+    ggml_cann_release_resources(ctx, acl_scalar);
+}
+
+/**
+ * @brief Get or expand a cached float32 tensor filled with a scalar value.
+ *
+ * This function manages cached device memory for float32 tensors. If the current
+ * cache size is insufficient for the requested tensor shape, the old memory will
+ * be released and new memory will be allocated. The allocated buffer is then
+ * initialized either with zeros (when @p value == 0.0f) or with the given scalar
+ * value using CANN operations. Finally, an aclTensor object is created from the
+ * cached memory and returned.
+ *
+ * @param ctx           The CANN backend context that manages device memory.
+ * @param buffer        A pointer to the cached device buffer (will be allocated
+ *                      or reallocated if necessary).
+ * @param cache_element The current number of cached elements. This will be
+ *                      updated when the cache is expanded.
+ * @param ne            The tensor shape array (number of elements in each dimension).
+ * @param nb            The stride size for each dimension.
+ * @param dims          The number of tensor dimensions.
+ * @param value         The scalar value used to fill the tensor (supports zero
+ *                      initialization via memset or arbitrary values via fill_scalar).
+ * @return              An aclTensor pointer created from the cached buffer.
+ */
+static aclTensor* get_f32_cache_acl_tensor(
+    ggml_backend_cann_context& ctx,
+    void** buffer,
+    int64_t &cache_element,
+    int64_t* ne,
+    size_t* nb,
+    int64_t dims,
+    float value) {
+    // Calculate total number of elements
+    int64_t n_element = 1;
+    for (int i = 0; i < dims; i++) {
+        n_element *= ne[i];
+    }
+    size_t size = n_element * sizeof(float);
+
+    // Allocate or expand cache if needed
+    if (cache_element < n_element) {
+        if (*buffer != nullptr) {
+            aclrtFree(*buffer);
+            *buffer = nullptr;
+        }
+
+        ACL_CHECK(aclrtMalloc(buffer, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        cache_element = n_element;
+
+        // Initialize cache
+        if (value == 0.0f) {
+            ACL_CHECK(aclrtMemsetAsync(*buffer, size, 0, size, ctx.stream()));
+        } else {
+            int64_t pool_ne[1] = { n_element };
+            size_t pool_nb[1] = { sizeof(float) };
+            aclTensor* acl_value = ggml_cann_create_tensor(
+                *buffer, ACL_FLOAT, sizeof(float), pool_ne, pool_nb, 1);
+            aclnn_fill_scalar(ctx, 1, acl_value);
+            ggml_cann_release_resources(ctx, acl_value);
+        }
+    }
+
+    return ggml_cann_create_tensor(*buffer, ACL_FLOAT, sizeof(float), ne, nb, dims);
+}
+
 void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
 
@@ -875,20 +955,39 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
-    size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src);
-    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
 
-    aclTensor* acl_gamma = aclnn_values(
-        ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, 1,
-        ggml_cann_type_mapping(src->type), ggml_element_size(src));
+    // build gamma, one...
+    size_t acl_gamma_nb[GGML_MAX_DIMS];
+    acl_gamma_nb[0] = sizeof(float);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        acl_gamma_nb[i] = acl_gamma_nb[i - 1] * src->ne[i - 1];
+    }
+    aclTensor* acl_gamma = get_f32_cache_acl_tensor(
+        ctx,
+        &ctx.f32_one_cache,
+        ctx.f32_one_cache_element,
+        src->ne,
+        acl_gamma_nb,
+        1,        // dims
+        1.0f      // value
+    );
+
+    // build rstd, zero...
+    size_t acl_rstd_nb[GGML_MAX_DIMS];
+    acl_rstd_nb[0] = sizeof(float);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        acl_rstd_nb[i] = acl_rstd_nb[i - 1] * src->ne[i - 1];
+    }
+    aclTensor* acl_rstd = get_f32_cache_acl_tensor(
+        ctx,
+        &ctx.f32_zero_cache,
+        ctx.f32_zero_cache_element,
+        src->ne,
+        acl_rstd_nb,
+        GGML_MAX_DIMS,
+        0.0f      // value
+    );
 
-    size_t zero_tensor_n_bytes =
-        src->ne[1] * src->ne[2] * src->ne[3] * ggml_element_size(src);
-    ggml_cann_pool_alloc zero_tensor_allocator(ctx.pool(), zero_tensor_n_bytes);
-    aclTensor* acl_rstd =
-        aclnn_zero(ctx, zero_tensor_allocator.get(), zero_tensor_n_bytes,
-                   src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
-                   ggml_element_size(src));
     GGML_CANN_CALL_ACLNN_OP(ctx, RmsNorm, acl_src, acl_gamma, eps, acl_dst, acl_rstd);
     ggml_cann_release_resources(ctx, acl_src, acl_dst, acl_gamma, acl_rstd);
 }
@@ -903,14 +1002,13 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst,
 
     const int n_past = ((int32_t*)dst->op_params)[0];
 
-    size_t one_tensor_n_bytes = src->ne[0] * src->ne[1] * src->ne[2] *
-                                src->ne[3] * ggml_element_size(src);
-    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
+    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), ggml_nbytes(src));
+    void* buffer = one_tensor_allocator.get();
 
-    aclTensor* mask_tensor =
-        aclnn_values(ctx, one_tensor_allocator.get(), one_tensor_n_bytes,
-                     src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
-                     ggml_element_size(src), value);
+    aclTensor* mask_tensor = ggml_cann_create_tensor(buffer, ggml_cann_type_mapping(src->type),
+        ggml_type_size(src->type), src->ne, src->nb, GGML_MAX_DIMS);
+
+    aclnn_fill_scalar(ctx, value, mask_tensor);
 
     aclScalar* alpha = nullptr;
     float alphaValue = 1.0f;
@@ -1159,12 +1257,20 @@ static void aclnn_exp(ggml_backend_cann_context& ctx, aclTensor* acl_src) {
 
 void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
                       aclTensor* acl_dst) {
-    GGML_CANN_CALL_ACLNN_OP(ctx, Cos, acl_src, acl_dst);
+    if(acl_dst == nullptr) {
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceCos, acl_src);
+    } else {
+        GGML_CANN_CALL_ACLNN_OP(ctx, Cos, acl_src, acl_dst);
+    }
 }
 
 void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
                       aclTensor* acl_dst) {
-    GGML_CANN_CALL_ACLNN_OP(ctx, Sin, acl_src, acl_dst);
+    if(acl_dst == nullptr) {
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceSin, acl_src);
+    } else {
+        GGML_CANN_CALL_ACLNN_OP(ctx, Sin, acl_src, acl_dst);
+    }
 }
 
 void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx,
@@ -1277,23 +1383,6 @@ void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx,
         tmp_permute_tensor, tmp_mul_tensor, acl_dst);
 }
 
-/**
- * @brief Fills a tensor with a scalar value.
- *
- * This function fills the destination tensor `acl_dst` with the scalar value
- * `scalar`.
- *
- * @param ctx The context for the CANN backend operations.
- * @param scalar The scalar value used to fill the tensor.
- * @param acl_dst The destination tensor to be filled with the scalar value.
- */
-static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar,
-                              aclTensor* acl_dst) {
-    auto acl_scalar = aclCreateScalar(&scalar, aclDataType::ACL_FLOAT);
-    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceFillScalar, acl_dst, acl_scalar);
-    ggml_cann_release_resources(ctx, acl_scalar);
-}
-
 /**
  * @brief Raises each element of a tensor to the power of the corresponding
  * element in another tensor.
@@ -1338,17 +1427,17 @@ static void aclnn_pow_tensor_tensor(ggml_backend_cann_context& ctx,
 static void aclnn_get_slope_inner(ggml_backend_cann_context& ctx, void* slope_buffer,
     float m, int64_t size, float start, float stop, float step){
     int64_t ne[] = {size};
-    size_t nb[] = {sizeof(float)};
+    size_t nb[] = {sizeof(uint16_t)};
 
-    ggml_cann_pool_alloc arange_allocator(ctx.pool(), size * sizeof(float));
+    ggml_cann_pool_alloc arange_allocator(ctx.pool(), size * sizeof(uint16_t));
     void* arange_buffer = arange_allocator.get();
 
     aclTensor* arange_tensor = ggml_cann_create_tensor(
-        arange_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
+        arange_buffer, ACL_FLOAT16, sizeof(uint16_t), ne, nb, 1);
     aclnn_arange(ctx, arange_tensor, start, stop, step, size);
 
     aclTensor* slope_tensor = ggml_cann_create_tensor(
-        slope_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
+        slope_buffer, ACL_FLOAT16, sizeof(uint16_t), ne, nb, 1);
 
     aclScalar* sc = aclCreateScalar(&m, aclDataType::ACL_FLOAT);
 
@@ -2140,13 +2229,54 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx,
     ggml_cann_release_resources(ctx, acl_index, acl_value);
 }
 
+/**
+ * @brief Initializes and caches sine/cosine positional encoding values
+ *        (used in RoPE, Rotary Position Embedding) for attention layers.
+ *
+ * This function computes and caches the sin/cos values of
+ * θ = position * theta_scale for RoPE encoding. The cache is shared
+ * across attention layers, and only the first attention layer will
+ * trigger initialization. The cache includes repeated sin/cos values
+ * with different repeat methods depending on the @param is_neox flag.
+ *
+ * Steps performed by this function:
+ *   1. Identify whether the target tensor belongs to Q/K in attention
+ *      and restrict computation to the first layer only.
+ *   2. Initialize the theta scale array (arange → power → freq scaling).
+ *   3. Allocate sin/cos caches if the max prompt length increases.
+ *   4. Compute θ = position * theta_scale.
+ *   5. Compute sin(θ), cos(θ) and optionally scale by attn_factor.
+ *   6. Expand sin/cos values by repeat or repeat_interleave depending
+ *      on whether @param is_neox is enabled.
+ *   7. Store the computed values into persistent buffers
+ *      (ctx.rope_sin_ptr / ctx.rope_cos_ptr).
+ *
+ * @param ctx         The CANN backend context, holding memory pool,
+ *                    stream, and persistent buffers for rope init/cache.
+ * @param dst         The destination ggml_tensor whose computation
+ *                    depends on the cached RoPE values (usually Qcur/Kcur).
+ * @param theta_scale Scalar exponent base for computing theta scale values.
+ * @param freq_scale  Frequency scaling factor, applied to theta scale.
+ * @param attn_factor Attention scaling factor, applied to sin/cos.
+ * @param is_neox     Whether to use Neox-style repeat strategy
+ *                    (dim expansion vs repeat_interleave).
+ */
 static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
-                             aclTensor* acl_cos_repeat_tensor,
-                             aclTensor* acl_sin_repeat_tensor,
                              float theta_scale, float freq_scale,
                              float attn_factor, bool is_neox) {
     // int sin/cos cache, cache has different repeat method depond on
     // @param.is_neox
+    bool is_q = (std::strncmp(dst->name, "Qcur-", 5) == 0);
+    bool is_k = (std::strncmp(dst->name, "Kcur-", 5) == 0);
+
+    // used for accuracy testing
+    bool is_attention = is_q || is_k;
+
+    // just compute in first layer in attention
+    bool is_fisrt_layer = (std::strncmp(dst->name, "Qcur-0", GGML_MAX_NAME) == 0);
+    if(is_attention && !is_fisrt_layer) {
+        return;
+    }
 
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src1 = dst->src[1];  // position
@@ -2172,21 +2302,16 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         theta_nb[i] = theta_nb[i - 1] * theta_ne[i - 1];
     }
 
-    bool is_q = (std::strncmp(dst->name, "Qcur-", 5) == 0);
-    bool is_k = (std::strncmp(dst->name, "Kcur-", 5) == 0);
-
-    // used for accuracy testing
-    bool is_attention = is_q || is_k;
-
-    if(ctx.init_ptr == nullptr || !is_attention) {
+    // init theta scale, just one time
+    if(ctx.rope_init_ptr == nullptr || !is_attention) {
         // theta_scale arange, [0,1,...,ne00/2 - 1]
-        if(ctx.init_ptr != nullptr){
-            ACL_CHECK(aclrtFree(ctx.init_ptr));
+        if(ctx.rope_init_ptr != nullptr){
+            ACL_CHECK(aclrtFree(ctx.rope_init_ptr));
         }
-        ACL_CHECK(aclrtMalloc(&ctx.init_ptr, theta_scale_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+        ACL_CHECK(aclrtMalloc(&ctx.rope_init_ptr, theta_scale_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
 
         aclTensor* acl_theta_scale_tensor =
-            ggml_cann_create_tensor(ctx.init_ptr, ACL_FLOAT, sizeof(float_t),
+            ggml_cann_create_tensor(ctx.rope_init_ptr, ACL_FLOAT, sizeof(float_t),
                                     theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
         float start = 0;
         float step = 1;
@@ -2216,67 +2341,55 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         ggml_cann_release_resources(ctx, acl_theta_scale_tensor,acl_theta_scale);
     }
 
-    if(ctx.sin_ptr == nullptr) {
-        int64_t theta_length = theta_scale_length * ctx.max_prompt_length;
-        ACL_CHECK(aclrtMalloc(&ctx.sin_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
-        ACL_CHECK(aclrtMalloc(&ctx.cos_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
-    }
+    // init sin_repeat && cos_repeat, one token just init in 0 layer
     if(position_length > ctx.max_prompt_length) {
         ctx.max_prompt_length = position_length;
-        int64_t theta_length = theta_scale_length * ctx.max_prompt_length;
-        ACL_CHECK(aclrtFree(ctx.sin_ptr));
-        ACL_CHECK(aclrtFree(ctx.cos_ptr));
-        ACL_CHECK(aclrtMalloc(&ctx.sin_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
-        ACL_CHECK(aclrtMalloc(&ctx.cos_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+        int64_t repeat_theta_length = theta_scale_length * ctx.max_prompt_length * 2;
+        if(ctx.rope_sin_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(ctx.rope_sin_ptr));
+            ACL_CHECK(aclrtFree(ctx.rope_cos_ptr));
+        }
+        ACL_CHECK(aclrtMalloc(&ctx.rope_sin_ptr, repeat_theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+        ACL_CHECK(aclrtMalloc(&ctx.rope_cos_ptr, repeat_theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
     }
 
-    bool is_fisrt_layer = (std::strncmp(dst->name, "Qcur-0", GGML_MAX_NAME) == 0);
-
-    if(is_fisrt_layer || !is_attention) {
-
-        aclTensor* acl_theta_scale_tensor =
-            ggml_cann_create_tensor(ctx.init_ptr, ACL_FLOAT, sizeof(float_t),
+    aclTensor* acl_theta_scale_tensor =
+            ggml_cann_create_tensor(ctx.rope_init_ptr, ACL_FLOAT, sizeof(float_t),
                                     theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
 
-        // position
-        aclTensor* acl_position_tensor = ggml_cann_create_tensor(
-            src1->data, ggml_cann_type_mapping(src1->type),
-            ggml_type_size(src1->type), position_ne, position_nb, GGML_MAX_DIMS);
+    // position
+    aclTensor* acl_position_tensor = ggml_cann_create_tensor(
+        src1->data, ggml_cann_type_mapping(src1->type),
+        ggml_type_size(src1->type), position_ne, position_nb, GGML_MAX_DIMS);
 
-        // power * position
-        int64_t theta_length = theta_scale_length * position_length;
-        ggml_cann_pool_alloc theta_allocator(ctx.pool(),
-                                            theta_length * sizeof(float_t));
-        void* theta_buffer = theta_allocator.get();
+    // power * position
+    int64_t theta_length = theta_scale_length * position_length;
+    ggml_cann_pool_alloc theta_allocator(ctx.pool(),
+                                        theta_length * sizeof(float_t));
+    void* theta_buffer = theta_allocator.get();
 
-        aclTensor* acl_theta_tensor =
-            ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float_t),
-                                    theta_ne, theta_nb, GGML_MAX_DIMS);
-        aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
-                acl_theta_tensor);
-
-        // sin/cos
-        aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
-            ctx.sin_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
-            GGML_MAX_DIMS, ACL_FORMAT_ND);
-        aclnn_sin(ctx, acl_theta_tensor, acl_sin_tensor);
-
-        aclTensor* acl_cos_tensor = ggml_cann_create_tensor(
-            ctx.cos_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
-            GGML_MAX_DIMS, ACL_FORMAT_ND);
-        aclnn_cos(ctx, acl_theta_tensor, acl_cos_tensor);
-
-        // release
-        ggml_cann_release_resources(ctx, acl_theta_scale_tensor, acl_position_tensor,
-            acl_theta_tensor, acl_sin_tensor, acl_cos_tensor);
-    }
+    aclTensor* acl_theta_tensor =
+        ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float_t),
+                                theta_ne, theta_nb, GGML_MAX_DIMS);
+    aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
+            acl_theta_tensor);
 
+    // sin/cos
+    ggml_cann_pool_alloc sin_allocator(ctx.pool(),
+                                    theta_length * sizeof(float_t));
+    void* sin_buffer = sin_allocator.get();
     aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
-            ctx.sin_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
-            GGML_MAX_DIMS, ACL_FORMAT_ND);
+        sin_buffer, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
+        GGML_MAX_DIMS, ACL_FORMAT_ND);
+    aclnn_sin(ctx, acl_theta_tensor, acl_sin_tensor);
+
+    ggml_cann_pool_alloc cos_allocator(ctx.pool(),
+                                    theta_length * sizeof(float_t));
+    void* cos_buffer = cos_allocator.get();
     aclTensor* acl_cos_tensor = ggml_cann_create_tensor(
-            ctx.cos_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
-            GGML_MAX_DIMS, ACL_FORMAT_ND);
+        cos_buffer, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
+        GGML_MAX_DIMS, ACL_FORMAT_ND);
+    aclnn_cos(ctx, acl_theta_tensor, acl_cos_tensor);
 
     // attn_factor
     if (attn_factor != 1) {
@@ -2284,6 +2397,19 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         aclnn_muls(ctx, acl_cos_tensor, attn_factor, nullptr, true);
     }
 
+    int64_t sin_reshape_ne[4] = {ne00, 1, ne02, 1};
+    size_t sin_reshape_nb[GGML_MAX_DIMS];
+    sin_reshape_nb[0] = sizeof(float_t);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        sin_reshape_nb[i] = sin_reshape_nb[i - 1] * sin_reshape_ne[i - 1];
+    }
+    aclTensor* acl_sin_repeat_tensor =
+        ggml_cann_create_tensor(ctx.rope_sin_ptr, ACL_FLOAT, sizeof(float_t),
+                                sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
+    aclTensor* acl_cos_repeat_tensor =
+        ggml_cann_create_tensor(ctx.rope_cos_ptr, ACL_FLOAT, sizeof(float_t),
+                                sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
+
     // repeat
     if (is_neox) {
         int64_t repeatsArray[] = {1, 1, 1, 2};
@@ -2299,8 +2425,9 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
                                 num_repeats, output_size);
     }
 
-    // release
-    ggml_cann_release_resources(ctx, acl_sin_tensor, acl_cos_tensor);
+    ggml_cann_release_resources(ctx, acl_theta_scale_tensor, acl_position_tensor,
+        acl_theta_tensor, acl_sin_tensor, acl_sin_repeat_tensor, acl_cos_tensor,
+        acl_cos_repeat_tensor);
 }
 
 #ifdef __cplusplus
@@ -2354,13 +2481,8 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
-    // init cos/sin cache
-    ggml_cann_pool_alloc sin_allocator(
-        ctx.pool(), ne00 * ne02 * sizeof(float_t));
-    ggml_cann_pool_alloc cos_allocator(
-        ctx.pool(), ne00 * ne02 * sizeof(float_t));
-    void* sin_buffer = sin_allocator.get();
-    void* cos_buffer = cos_allocator.get();
+    // init ctx.rope_cos/rope_sin cache
+    aclnn_cache_init(ctx, dst, theta_scale, freq_scale, attn_factor, is_neox);
 
     int64_t sin_reshape_ne[4] = {ne00, 1, ne02, 1};
     size_t sin_reshape_nb[GGML_MAX_DIMS];
@@ -2369,13 +2491,11 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         sin_reshape_nb[i] = sin_reshape_nb[i - 1] * sin_reshape_ne[i - 1];
     }
     aclTensor* acl_sin_reshape_tensor =
-        ggml_cann_create_tensor(sin_buffer, ACL_FLOAT, sizeof(float_t),
+        ggml_cann_create_tensor(ctx.rope_sin_ptr, ACL_FLOAT, sizeof(float_t),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
     aclTensor* acl_cos_reshape_tensor =
-        ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float_t),
+        ggml_cann_create_tensor(ctx.rope_cos_ptr, ACL_FLOAT, sizeof(float_t),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
-    aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor,
-                    theta_scale, freq_scale, attn_factor, is_neox);
 
     aclTensor* acl_src = ggml_cann_create_tensor(src0);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
@@ -3060,11 +3180,38 @@ void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
 void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
 
-    ggml_tensor* src0 = dst->src[0]; // q, fp32
-    ggml_tensor* src1 = dst->src[1]; // k, fp16
-    ggml_tensor* src2 = dst->src[2]; // v, fp16
+    ggml_tensor* src0 = dst->src[0]; // q, fp32 | B, N, S, D (uncont) -> B, S, N, D (cont)
+    ggml_tensor* src1 = dst->src[1]; // k, fp16 | B, N, S, D (uncont) -> B, S, N, D (cont)
+    ggml_tensor* src2 = dst->src[2]; // v, fp16 | B, N, S, D (uncont) -> B, S, N, D (cont)
     ggml_tensor* src3 = dst->src[3]; // mask, fp16
 
+    // B, N, S, D (uncont) -> B, S, N, D (cont)
+    int64_t src0_bsnd_ne[GGML_MAX_DIMS];
+    memcpy(src0_bsnd_ne, src0->ne, GGML_MAX_DIMS * sizeof(int64_t));
+    size_t src0_bsnd_nb[GGML_MAX_DIMS];
+    memcpy(src0_bsnd_nb, src0->nb, GGML_MAX_DIMS * sizeof(size_t));
+    int64_t src1_bsnd_ne[GGML_MAX_DIMS];
+    memcpy(src1_bsnd_ne, src1->ne, GGML_MAX_DIMS * sizeof(int64_t));
+    size_t src1_bsnd_nb[GGML_MAX_DIMS];
+    memcpy(src1_bsnd_nb, src1->nb, GGML_MAX_DIMS * sizeof(size_t));
+    int64_t src2_bsnd_ne[GGML_MAX_DIMS];
+    memcpy(src2_bsnd_ne, src2->ne, GGML_MAX_DIMS * sizeof(int64_t));
+    size_t src2_bsnd_nb[GGML_MAX_DIMS];
+    memcpy(src2_bsnd_nb, src2->nb, GGML_MAX_DIMS * sizeof(size_t));
+
+    auto transpose12 = [](int64_t* ne, size_t* nb) {
+        int64_t ne_tmp = ne[1];
+        size_t  nb_tmp = nb[1];
+        ne[1] = ne[2];
+        nb[1] = nb[2];
+        ne[2] = ne_tmp;
+        nb[2] = nb_tmp;
+    };
+
+    transpose12(src0_bsnd_ne, src0_bsnd_nb);
+    transpose12(src1_bsnd_ne, src1_bsnd_nb);
+    transpose12(src2_bsnd_ne, src2_bsnd_nb);
+
     float maxBias = 0.0f;
     float scaleValue = 1.0f;
     float logitSoftcap = 0.0f;
@@ -3086,11 +3233,12 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
         void* src0_f16_buffer = nullptr;
 
         if(ggml_cann_type_mapping(src0->type) != faDataType){
-            aclTensor* acl_src0_f32_tensor = ggml_cann_create_tensor(src0);
+            aclTensor* acl_src0_f32_tensor = ggml_cann_create_tensor(src0, src0_bsnd_ne,
+                src0_bsnd_nb, GGML_MAX_DIMS);
             src0_f16_buffer = src0_f16_allocator.alloc(
                                     ggml_nelements(src0) * faElemSize);
 
-            int64_t* src0_f16_ne = src0->ne;
+            int64_t* src0_f16_ne = src0_bsnd_ne;
             size_t   src0_f16_nb[GGML_MAX_DIMS];
             src0_f16_nb[0] = sizeof(uint16_t);
             for(int i = 1; i < GGML_MAX_DIMS; ++i){
@@ -3104,20 +3252,23 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
             aclnn_cast(ctx, acl_src0_f32_tensor, acl_src0_f16_tensor, faDataType);
             ggml_cann_release_resources(ctx, acl_src0_f32_tensor);
         }else{
-            acl_src0_f16_tensor = ggml_cann_create_tensor(src0);
+            acl_src0_f16_tensor = ggml_cann_create_tensor(src0, src0_bsnd_ne,
+                src0_bsnd_nb, GGML_MAX_DIMS);
         }
 
         // Step 2: create the acl tensors for src1 (Key), src2 (Value),
         //         and the direct output from FusedInferAttention
 
-        acl_src1_f16_tensor = ggml_cann_create_tensor(src1);
-        acl_src2_f16_tensor = ggml_cann_create_tensor(src2);
+        acl_src1_f16_tensor = ggml_cann_create_tensor(src1, src1_bsnd_ne,
+            src1_bsnd_nb, GGML_MAX_DIMS);
+        acl_src2_f16_tensor = ggml_cann_create_tensor(src2, src2_bsnd_ne,
+            src2_bsnd_nb, GGML_MAX_DIMS);
 
         ggml_cann_pool_alloc out_f16_allocator(ctx.pool());
         void* out_f16_buffer = out_f16_allocator.alloc(
                                     ggml_nelements(dst) * faElemSize);
 
-        int64_t* out_f16_ne = src0->ne;
+        int64_t* out_f16_ne = src0_bsnd_ne;
         size_t out_f16_nb[GGML_MAX_DIMS];
         out_f16_nb[0] = faElemSize;
         for(int i = 1; i < GGML_MAX_DIMS; ++i){
@@ -3131,88 +3282,81 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
 
         // Step 3: create the PSEShift tensor if needed
         //         this tensor is considered as mask (f16) in the llama.cpp
-
         aclTensor* bcast_pse_tensor = nullptr;
-        int64_t bcast_pse_ne[GGML_MAX_DIMS];
-        size_t bcast_pse_nb[GGML_MAX_DIMS];
         ggml_cann_pool_alloc bcast_pse_allocator(ctx.pool());
-        void* bcast_pse_buffer = nullptr;
-
         if(src3 != nullptr){
-            bcast_pse_buffer = bcast_pse_allocator.alloc(
-                            ggml_nelements(src3) * src0->ne[2] * sizeof(uint16_t));
+            // Construct the truncated pse tensor (common for prefill/decode)
+            int64_t trunc_pse_ne[GGML_MAX_DIMS] = {
+                src3->ne[0],        // D
+                src0->ne[1],        // S (number of Q tokens)
+                src3->ne[2],        // mask N
+                src3->ne[3]         // B
+            };
+            size_t* trunc_pse_nb = src3->nb;
 
-            if(src0->ne[1] > 1){
-                // Case 1: broadcast pse for prefill stage with multiple head
-                aclTensor* acl_mask_f16_tensor = ggml_cann_create_tensor(src3);
-                bcast_pse_ne[0] = src3->ne[0];
-                bcast_pse_ne[1] = src3->ne[1];
-                bcast_pse_ne[2] = src0->ne[2];
-                bcast_pse_ne[3] = src3->ne[3];
+            aclTensor* acl_mask_f16_trunc_tensor = ggml_cann_create_tensor(
+                src3->data, ACL_FLOAT16, sizeof(uint16_t),
+                trunc_pse_ne, trunc_pse_nb, GGML_MAX_DIMS
+            );
 
+            int64_t bcast_pse_ne[GGML_MAX_DIMS];
+            size_t bcast_pse_nb[GGML_MAX_DIMS];
+            bcast_pse_ne[0] = src3->ne[0];      // D
+            bcast_pse_ne[1] = src0->ne[1];      // S
+            bcast_pse_ne[2] = src0->ne[2];      // N (num_heads)
+            bcast_pse_ne[3] = src3->ne[3];      // B
+            if (maxBias == 0.0f) {
+                // When maxBias == 0.0f, use nb = 0 reduce once repeat (Qwen2)
+                // Construct the bcast tensor (simulate repeat on the head dimension using stride=0)
                 bcast_pse_nb[0] = sizeof(uint16_t);
-                for(int i = 1; i < GGML_MAX_DIMS; ++i){
-                    bcast_pse_nb[i] = bcast_pse_nb[i - 1] * bcast_pse_ne[i - 1];
-                }
+                bcast_pse_nb[1] = bcast_pse_nb[0] * bcast_pse_ne[0];
+                bcast_pse_nb[2] = 0;                // <---- the head dimension shares the same data
+                bcast_pse_nb[3] = src3->nb[3];
 
                 bcast_pse_tensor = ggml_cann_create_tensor(
-                    bcast_pse_buffer, ACL_FLOAT16, sizeof(uint16_t),
-                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS);
-
-                int64_t repeats[] = {1, src0->ne[2], 1, 1};
-                aclnn_repeat(ctx, acl_mask_f16_tensor, bcast_pse_tensor, repeats);
-
-                ggml_cann_release_resources(ctx, acl_mask_f16_tensor);
-            }else{
-                // Case 2: trunc the first row and broadcast pse for decode stage with multiple head
-                int64_t trunc_pse_ne[GGML_MAX_DIMS] = {src3->ne[0], src0->ne[1], src3->ne[2], src3->ne[3]};
-                size_t* trunc_pse_nb = src3->nb;
-
-                aclTensor* acl_mask_f16_trunc_tensor = ggml_cann_create_tensor(
                     src3->data, ACL_FLOAT16, sizeof(uint16_t),
-                    trunc_pse_ne, trunc_pse_nb, GGML_MAX_DIMS);
-
-                bcast_pse_ne[0] = src3->ne[0];
-                bcast_pse_ne[1] = src0->ne[1];
-                bcast_pse_ne[2] = src0->ne[2];
-                bcast_pse_ne[3] = src3->ne[3];
+                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS
+                );
 
+                ggml_cann_release_resources(ctx, acl_mask_f16_trunc_tensor);
+            } else {
                 bcast_pse_nb[0] = sizeof(uint16_t);
-                for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                for (int i = 1; i < GGML_MAX_DIMS; i++) {
                     bcast_pse_nb[i] = bcast_pse_nb[i - 1] * bcast_pse_ne[i - 1];
                 }
 
+                void* bcast_pse_buffer = bcast_pse_allocator.alloc(
+                    ggml_nelements(src3) * src0->ne[2] * sizeof(uint16_t)
+                );
+
                 bcast_pse_tensor = ggml_cann_create_tensor(
                     bcast_pse_buffer, ACL_FLOAT16, sizeof(uint16_t),
-                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS);
+                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS
+                );
 
                 int64_t repeats[] = {1, src0->ne[2], 1, 1};
                 aclnn_repeat(ctx, acl_mask_f16_trunc_tensor, bcast_pse_tensor, repeats);
 
-                ggml_cann_release_resources(ctx, acl_mask_f16_trunc_tensor);
-            }
-
-            // Compute the slope if needed. Derived from ggml_cann_softmax().
-            if(maxBias != 0.0f){
                 // alibi
+                // Compute the slope if needed. Derived from ggml_cann_softmax().
                 const int64_t n_heads = src0->ne[2];
-                ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(float));
+                ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(uint16_t));
                 void* slope_buffer = slope_allocator.get();
                 aclnn_get_slope(ctx, n_heads, slope_buffer, maxBias);
 
                 int64_t slope_ne[] = {1, 1, n_heads, 1};
                 size_t slope_nb[GGML_MAX_DIMS];
-                slope_nb[0] = sizeof(float);
+                slope_nb[0] = sizeof(uint16_t);
                 for(int i = 1;i<GGML_MAX_DIMS;i++) {
                     slope_nb[i] = slope_nb[i-1] * slope_ne[0];
                 }
 
                 aclTensor* slope_tensor = ggml_cann_create_tensor(
-                    slope_buffer, ACL_FLOAT, sizeof(float),
+                    slope_buffer, ACL_FLOAT16, sizeof(uint16_t),
                     slope_ne, slope_nb, GGML_MAX_DIMS);
                 GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, bcast_pse_tensor, slope_tensor);
 
-                ggml_cann_release_resources(ctx, slope_tensor);
+                ggml_cann_release_resources(ctx, slope_tensor, acl_mask_f16_trunc_tensor);
             }
         }
 
@@ -3229,7 +3373,7 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
         // double  scaleValue = 1 / sqrt(src0->ne[0]); // 1/sqrt(d)
         int64_t preTokens = 65535;
         int64_t nextTokens = 65535;
-        char layout[5] = {'B', 'N', 'S', 'D', 0};
+        char layout[5] = {'B', 'S', 'N', 'D', 0};
         int64_t sparseMode = 0;
         int64_t innerPrecise = (src0->ne[1] == 1) ? 0 : 2;
         int64_t blockSize = 0;
@@ -3266,32 +3410,9 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
         );
 
         // Step 6: post-processing, permute and cast to f32
-
-        int64_t new_dim[] = {0, 2, 1, 3};
         aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
-
-        if(ggml_cann_type_mapping(dst->type) != faDataType){
-            ggml_cann_pool_alloc perm_out_f16_allocator(ctx.pool());
-            perm_out_f16_allocator.alloc(ggml_nelements(dst) * faElemSize);
-            void* perm_out_f16_buffer = perm_out_f16_allocator.get();
-
-            int64_t* perm_out_f16_ne = dst->ne;
-            size_t  perm_out_f16_nb[GGML_MAX_DIMS];
-            perm_out_f16_nb[0] = faElemSize;
-            for(int i = 1; i < GGML_MAX_DIMS; ++i){
-                perm_out_f16_nb[i] = perm_out_f16_nb[i - 1] * perm_out_f16_ne[i - 1];
-            }
-            aclTensor* acl_perm_out_f16_tensor = ggml_cann_create_tensor(
-                perm_out_f16_buffer, faDataType, faElemSize,
-                perm_out_f16_ne, perm_out_f16_nb, GGML_MAX_DIMS);
-            aclnn_permute(ctx, acl_dst_f16_tensor, acl_perm_out_f16_tensor, new_dim, GGML_MAX_DIMS);
-            aclnn_cast(ctx,
-                acl_perm_out_f16_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
-            ggml_cann_release_resources(ctx, acl_perm_out_f16_tensor);
-        }else{
-            // only need to permute
-            aclnn_permute(ctx, acl_dst_f16_tensor, acl_dst_tensor, new_dim, GGML_MAX_DIMS);
-        }
+        // TODO: when dst is fp16, don't need cast
+        aclnn_cast(ctx, acl_dst_f16_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
         ggml_cann_release_resources(ctx, acl_src0_f16_tensor,
                                          acl_src1_f16_tensor,
                                          acl_src2_f16_tensor,

ggml/src/ggml-cann/common.h

@@ -368,10 +368,6 @@ struct ggml_backend_cann_context {
     std::string name;                /**< Name of the device. */
     std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
-    void* init_ptr = nullptr;
-    void* sin_ptr = nullptr;
-    void* cos_ptr = nullptr;
-    int64_t max_prompt_length = 65536;
 #ifdef USE_ACL_GRAPH
     /// Cached CANN ACL graph used for executing the current ggml computation graph.
     std::unique_ptr<ggml_cann_graph> cann_graph;
@@ -379,6 +375,16 @@ struct ggml_backend_cann_context {
     cann_task_queue task_queue;
     bool async_mode;
     bool support_set_rows;
+    // Rope Cache
+    void* rope_init_ptr = nullptr;
+    void* rope_sin_ptr = nullptr;
+    void* rope_cos_ptr = nullptr;
+    int64_t max_prompt_length = 0;
+    // Constant Pool
+    void* f32_zero_cache = nullptr;
+    void* f32_one_cache = nullptr;
+    int64_t f32_zero_cache_element = 0;
+    int64_t f32_one_cache_element = 0;
 
     aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
@@ -418,14 +424,20 @@ struct ggml_backend_cann_context {
                 ACL_CHECK(aclrtDestroyStream(streams[i]));
             }
         }
-        if(init_ptr != nullptr) {
-            ACL_CHECK(aclrtFree(init_ptr));
+        if(rope_init_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(rope_init_ptr));
         }
-        if(sin_ptr != nullptr) {
-            ACL_CHECK(aclrtFree(sin_ptr));
+        if(rope_sin_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(rope_sin_ptr));
         }
-        if(cos_ptr != nullptr) {
-            ACL_CHECK(aclrtFree(cos_ptr));
+        if(rope_cos_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(rope_cos_ptr));
+        }
+        if(f32_zero_cache != nullptr) {
+            ACL_CHECK(aclrtFree(f32_zero_cache));
+        }
+        if(f32_one_cache != nullptr) {
+            ACL_CHECK(aclrtFree(f32_one_cache));
         }
     }
 

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2336,7 +2336,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
 #ifdef ASCEND_310P
-                    // Q4 && Q8 per group is not suppor on 310p device
+                    // Q4 && Q8 per group is not support on 310p device
                     return false;
 #endif
                     // only support contiguous for quantized types.
@@ -2354,7 +2354,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
 #ifdef ASCEND_310P
-                    // Q4 && Q8 per group is not suppor on 310p device
+                    // Q4 && Q8 per group is not support on 310p device
                     return false;
 #endif
                     // only support contiguous for quantized types.
@@ -2505,6 +2505,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             }
             return true;
         case GGML_OP_FLASH_ATTN_EXT:{
+#ifdef ASCEND_310P
+            // FA not support on 310p device
+            return false;
+#endif
             // derived from [ggml-cuda.cu]
             if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
                 return false;
@@ -2530,6 +2534,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // DeepSeek MLA
                 return false;
             }
+            if (op->src[0]->ne[0] % 16 != 0) {
+                // TODO: padding to support
+                return false;
+            }
             float logitSoftcap = 0.0f;
             memcpy(&logitSoftcap,  (float*)op->op_params + 2, sizeof(float));
             if(logitSoftcap != 0.0f) {

ggml/src/ggml-cpu/CMakeLists.txt

@@ -435,7 +435,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             )
         if (GGML_RVV)
             if (GGML_XTHEADVECTOR)
-                list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
+                list(APPEND ARCH_FLAGS -march=rv64gc_zfhmin_xtheadvector -mabi=lp64d)
             elseif (GGML_RV_ZFH)
                 list(APPEND ARCH_FLAGS -march=rv64gcv_zfhmin -mabi=lp64d)
             else()

ggml/src/ggml-cpu/arch-fallback.h

@@ -150,8 +150,6 @@
 #elif defined(__s390x__)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
-#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K

ggml/src/ggml-cpu/arch/s390/quants.c

@@ -23,6 +23,27 @@
 
 #define UNUSED GGML_UNUSED
 
+#if defined(__VXE__) || defined(__VXE2__)
+#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
+#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
+#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
+#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
+#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
+#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
+#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
+#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
+
+// precomputed tables for expanding 8bits to 8 bytes:
+static const __attribute__((aligned(16))) uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b ) << 4
+static const __attribute__((aligned(16))) uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
+
+// permute mask for byteswapping
+static const uint8x16_t v_kperm = (const uint8x16_t){
+     7,  6,  5,  4,  3,  2, 1, 0,
+    15, 14, 13, 12, 11, 10, 9, 8
+};
+#endif
+
 void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
@@ -241,6 +262,301 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }
 
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0.0f;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    float32x4_t v_sum0 = vec_splats(0.0f);
+    float32x4_t v_sum1 = vec_splats(0.0f);
+
+    uint32_t qh0, qh1;
+    uint64_t tmp0[4], tmp1[4];
+
+    const uint8x16_t v_m = vec_splats((uint8_t)0x0F);
+
+    #pragma GCC unroll 4
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_0 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q5_0 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
+
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];
+
+        int8x16_t v_qh0l = vec_xl(0, (const int8_t *)(tmp0 + 0));
+        int8x16_t v_qh0h = vec_xl(0, (const int8_t *)(tmp0 + 2));
+        int8x16_t v_qh1l = vec_xl(0, (const int8_t *)(tmp1 + 0));
+        int8x16_t v_qh1h = vec_xl(0, (const int8_t *)(tmp1 + 2));
+
+        // required for fixing the byteorder
+        v_qh0l = vec_perm(v_qh0l, v_qh0l, v_kperm);
+        v_qh0h = vec_perm(v_qh0h, v_qh0h, v_kperm);
+        v_qh1l = vec_perm(v_qh1l, v_qh1l, v_kperm);
+        v_qh1h = vec_perm(v_qh1h, v_qh1h, v_kperm);
+
+        const uint8x16_t v_x0 = vec_xl(0, (const uint8_t *)x0->qs);
+        const uint8x16_t v_x1 = vec_xl(0, (const uint8_t *)x1->qs);
+
+        int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+        int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+        int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+        int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+        const int8x16_t v_x0lf = vec_sub(v_x0l, v_qh0l);
+        const int8x16_t v_x0hf = vec_sub(v_x0h, v_qh0h);
+        const int8x16_t v_x1lf = vec_sub(v_x1l, v_qh1l);
+        const int8x16_t v_x1hf = vec_sub(v_x1h, v_qh1h);
+
+        const int8x16_t v_y0l = vec_xl(0,       (const int8_t *)y0->qs);
+        const int8x16_t v_y0h = vec_xl(QK8_0/2, (const int8_t *)y0->qs);
+        const int8x16_t v_y1l = vec_xl(0,       (const int8_t *)y1->qs);
+        const int8x16_t v_y1h = vec_xl(QK8_0/2, (const int8_t *)y1->qs);
+
+        const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0lf, v_y0l), v_x0hf, v_y0h);
+        const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1lf, v_y1l), v_x1hf, v_y1h);
+
+        const float32x4_t v_xy0f = vec_float(v_xy0);
+        const float32x4_t v_xy1f = vec_float(v_xy1);
+
+        const float32x4_t v_d0 = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_d1 = vec_splats(GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d));
+
+        v_sum0 = vec_madd(v_xy0f, v_d0, v_sum0);
+        v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1);
+    }
+
+    sumf += vec_hsum(v_sum0) + vec_hsum(v_sum1);
+
+    #pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        const block_q5_0 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib];
+
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        uint64_t tmp[4];
+        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_1[(qh >> 24)       ];
+
+        int8x16_t v_qhl = vec_xl(0, (const int8_t *)(tmp + 0));
+        int8x16_t v_qhh = vec_xl(0, (const int8_t *)(tmp + 2));
+
+        // required for fixing the byteorder
+        v_qhl = vec_perm(v_qhl, v_qhl, v_kperm);
+        v_qhh = vec_perm(v_qhh, v_qhh, v_kperm);
+
+        const uint8x16_t v_x = vec_xl(0, (const uint8_t *)x0->qs);
+        int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        const int8x16_t v_xlf = vec_sub(v_xl, v_qhl);
+        const int8x16_t v_xhf = vec_sub(v_xh, v_qhh);
+
+        const int8x16_t v_yl = vec_xl(0,       (const int8_t *)y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, (const int8_t *)y0->qs);
+
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xlf, v_yl), v_xhf, v_yh);
+        const float32x4_t v_xyf = vec_float(v_xy);
+
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_acc = vec_madd(v_xyf, v_d, vec_splats(0.0f));
+
+        sumf += vec_hsum(v_acc);
+    }
+
+    *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0.0f;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    float32x4_t v_sum0 = vec_splats(0.0f);
+    float32x4_t v_sum1 = vec_splats(0.0f);
+
+    float summs0 = 0.0f;
+    float summs1 = 0.0f;
+
+    uint32_t qh0;
+    uint32_t qh1;
+
+    uint64_t tmp0[4];
+    uint64_t tmp1[4];
+
+    const uint8x16_t v_m = vec_splats((uint8_t)0x0F);
+
+    #pragma GCC unroll 4
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q5_1 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1];
+
+        summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+        summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
+
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_0[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_0[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_0[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_0[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_0[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_0[(qh1 >> 24)       ];
+
+        int8x16_t v_qh0l = vec_xl(0, (const int8_t *)(tmp0 + 0));
+        int8x16_t v_qh0h = vec_xl(0, (const int8_t *)(tmp0 + 2));
+        int8x16_t v_qh1l = vec_xl(0, (const int8_t *)(tmp1 + 0));
+        int8x16_t v_qh1h = vec_xl(0, (const int8_t *)(tmp1 + 2));
+
+        // required for fixing the byteorder
+        v_qh0l = vec_perm(v_qh0l, v_qh0l, v_kperm);
+        v_qh0h = vec_perm(v_qh0h, v_qh0h, v_kperm);
+        v_qh1l = vec_perm(v_qh1l, v_qh1l, v_kperm);
+        v_qh1h = vec_perm(v_qh1h, v_qh1h, v_kperm);
+
+        const uint8x16_t v_x0 = vec_xl(0, x0->qs);
+        const uint8x16_t v_x1 = vec_xl(0, x1->qs);
+
+        const int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+        const int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+        const int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+        const int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+        const int8x16_t v_x0lf = vec_or(v_x0l, v_qh0l);
+        const int8x16_t v_x0hf = vec_or(v_x0h, v_qh0h);
+        const int8x16_t v_x1lf = vec_or(v_x1l, v_qh1l);
+        const int8x16_t v_x1hf = vec_or(v_x1h, v_qh1h);
+
+        const int8x16_t v_y0l = vec_xl(0      , y0->qs);
+        const int8x16_t v_y0h = vec_xl(QK8_1/2, y0->qs);
+        const int8x16_t v_y1l = vec_xl(0      , y1->qs);
+        const int8x16_t v_y1h = vec_xl(QK8_1/2, y1->qs);
+
+        const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0lf, v_y0l), v_x0hf, v_y0h);
+        const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1lf, v_y1l), v_x1hf, v_y1h);
+
+        const float32x4_t v_xy0f = vec_float(v_xy0);
+        const float32x4_t v_xy1f = vec_float(v_xy1);
+
+        const float32x4_t v_d0 = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_d1 = vec_splats(GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d));
+
+        v_sum0 = vec_madd(v_xy0f, v_d0, v_sum0);
+        v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1);
+    }
+
+    sumf += vec_hsum(v_sum0) + vec_hsum(v_sum1) + summs0 + summs1;
+
+    #pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib];
+
+        float summs = GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        uint64_t tmp[4];
+        tmp[0] = table_b2b_0[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_0[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_0[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_0[(qh >> 24)       ];
+
+        int8x16_t v_qhl = vec_xl(0, (const int8_t *)(tmp + 0));
+        int8x16_t v_qhh = vec_xl(0, (const int8_t *)(tmp + 2));
+
+        // required for fixing the byteorder
+        v_qhl = vec_perm(v_qhl, v_qhl, v_kperm);
+        v_qhh = vec_perm(v_qhh, v_qhh, v_kperm);
+
+        const uint8x16_t v_x = vec_xl(0, x0->qs);
+        const int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        const int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        const int8x16_t v_xlf = vec_or(v_xl, v_qhl);
+        const int8x16_t v_xhf = vec_or(v_xh, v_qhh);
+
+        const int8x16_t v_yl = vec_xl(0      , y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_1/2, y0->qs);
+
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xlf, v_yl), v_xhf, v_yh);
+        const float32x4_t v_xyf = vec_float(v_xy);
+
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_acc = vec_madd(v_xyf, v_d, v_acc);
+
+        sumf += vec_hsum(v_acc) + summs;
+    }
+
+    *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

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

@@ -486,6 +486,14 @@ inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
     return v_abo + v_abe;
 }
 
+/**
+ * @see https://github.com/ggml-org/llama.cpp/pull/14037
+ */
+inline float vec_hsum(float32x4_t v) {
+    float32x4_t v_temp = v + vec_reve(v);
+    return v_temp[0] + v_temp[1];
+}
+
 inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
     const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
     return acc + (vec_unpackh(p) + vec_unpackl(p));

ggml/src/ggml-cpu/ggml-cpu.c

@@ -1880,6 +1880,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_conv_2d(params, tensor);
             } break;
+        case GGML_OP_CONV_3D:
+            {
+                ggml_compute_forward_conv_3d(params, tensor);
+            } break;
         case GGML_OP_CONV_2D_DW:
             {
                 ggml_compute_forward_conv_2d_dw(params, tensor);
@@ -2252,6 +2256,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_BACK:
         case GGML_OP_CONV_2D:
+        case GGML_OP_CONV_3D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
@@ -2773,6 +2778,7 @@ struct ggml_cplan ggml_graph_plan(
                         }
                     } break;
                 case GGML_OP_CONV_2D:
+                case GGML_OP_CONV_3D:
                     {
                         cur = GGML_IM2COL_WORK_SIZE;
                     } break;

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -2169,94 +2169,117 @@ class tinyBLAS_Q0_PPC {
 class tinyBLAS_PPC {
   public:
     tinyBLAS_PPC(int64_t k,
-                const float *A, int64_t lda,
-                const float *B, int64_t ldb,
-                float *C, int64_t ldc,
+                const float * A, int64_t lda,
+                const float * B, int64_t ldb,
+                float * C, int64_t ldc,
                 int ith, int nth)
         : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
     }
 
     void matmul(int64_t m, int64_t n) {
-       mnpack(0, m, 0, n);
+        int64_t mc = 256; int64_t nc = 256; int64_t kc = 256;
+        if (m % mc == 0 && n % nc == 0 && k % kc == 0) {
+            matmul_tiled(m, n, mc, nc, kc);
+        } else {
+            mnpack(0, m, 0, n);
+        }
     }
 
   private:
 
-    void (tinyBLAS_PPC::*kernel)(int64_t, int64_t);
-
-    inline void vector_permute_store_4(vector float *src, float *vecOffset) {
-       vector float t1, t2, t3, t4, t5, t6, t7, t8;
-           t1 = vec_mergeh(src[0], src[1]);
-           t2 = vec_mergeh(src[2], src[3]);
-           t3 = vec_mergel(src[0], src[1]);
-           t4 = vec_mergel(src[2], src[3]);
-
-           t5 = vec_xxpermdi(t1, t2, 0);
-           t6 = vec_xxpermdi(t1, t2, 3);
-           t7 = vec_xxpermdi(t3, t4, 0);
-           t8 = vec_xxpermdi(t3, t4, 3);
-
-           vec_xst(t5, 0, vecOffset);
-           vec_xst(t6, 0, vecOffset + 4);
-           vec_xst(t7, 0, vecOffset + 8);
-           vec_xst(t8, 0, vecOffset + 12);
-       }
-
-    inline void vector_permute_store_8(vector float *src, float *vecOffset) {
-       vector float t1, t2, t3, t4, t5, t6, t7, t8;
-          t1 = vec_mergeh(src[0], src[1]);
-          t2 = vec_mergeh(src[2], src[3]);
-          t3 = vec_mergeh(src[4], src[5]);
-          t4 = vec_mergeh(src[6], src[7]);
-
-          t5 = vec_xxpermdi(t1, t2, 0);
-          t6 = vec_xxpermdi(t3, t4, 0);
-          t7 = vec_xxpermdi(t1, t2, 3);
-          t8 = vec_xxpermdi(t3, t4, 3);
-
-          vec_xst(t5, 0, vecOffset);
-          vec_xst(t6, 0, vecOffset + 4);
-          vec_xst(t7, 0, vecOffset + 8);
-          vec_xst(t8, 0, vecOffset + 12);
-
-          t1 = vec_mergel(src[0], src[1]);
-          t2 = vec_mergel(src[2], src[3]);
-          t3 = vec_mergel(src[4], src[5]);
-          t4 = vec_mergel(src[6], src[7]);
-
-          t5 = vec_xxpermdi(t1, t2, 0);
-          t6 = vec_xxpermdi(t3, t4, 0);
-          t7 = vec_xxpermdi(t1, t2, 3);
-          t8 = vec_xxpermdi(t3, t4, 3);
-
-          vec_xst(t5, 0, vecOffset + 16);
-          vec_xst(t6, 0, vecOffset + 20);
-          vec_xst(t7, 0, vecOffset + 24);
-          vec_xst(t8, 0, vecOffset + 28);
+    inline void save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                *((float *)(C+ii+((jj+J)*ldc)+I)) = *((float *)&vec_C[I]+J);
+            }
+        }
     }
 
- void packTranspose(const float* a, int64_t lda, int rows, int cols, float* vec) {
+    inline void add_save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                float * c_ptr = (float *)(C+ii+((jj+J)*ldc)+I);
+                *c_ptr += *((float *)&vec_C[I]+J);
+            }
+        }
+    }
+
+    inline void vector_permute_store_4(vector float * src, float * vecOffset) {
+        vector float t1, t2, t3, t4, t5, t6, t7, t8;
+        t1 = vec_mergeh(src[0], src[1]);
+        t2 = vec_mergeh(src[2], src[3]);
+        t3 = vec_mergel(src[0], src[1]);
+        t4 = vec_mergel(src[2], src[3]);
+
+        t5 = vec_xxpermdi(t1, t2, 0);
+        t6 = vec_xxpermdi(t1, t2, 3);
+        t7 = vec_xxpermdi(t3, t4, 0);
+        t8 = vec_xxpermdi(t3, t4, 3);
+
+        vec_xst(t5, 0, vecOffset);
+        vec_xst(t6, 0, vecOffset + 4);
+        vec_xst(t7, 0, vecOffset + 8);
+        vec_xst(t8, 0, vecOffset + 12);
+    }
+
+    inline void vector_permute_store_8(vector float * src, float * vecOffset) {
+        vector float t1, t2, t3, t4, t5, t6, t7, t8;
+        t1 = vec_mergeh(src[0], src[1]);
+        t2 = vec_mergeh(src[2], src[3]);
+        t3 = vec_mergeh(src[4], src[5]);
+        t4 = vec_mergeh(src[6], src[7]);
+
+        t5 = vec_xxpermdi(t1, t2, 0);
+        t6 = vec_xxpermdi(t3, t4, 0);
+        t7 = vec_xxpermdi(t1, t2, 3);
+        t8 = vec_xxpermdi(t3, t4, 3);
+
+        vec_xst(t5, 0, vecOffset);
+        vec_xst(t6, 0, vecOffset + 4);
+        vec_xst(t7, 0, vecOffset + 8);
+        vec_xst(t8, 0, vecOffset + 12);
+
+        t1 = vec_mergel(src[0], src[1]);
+        t2 = vec_mergel(src[2], src[3]);
+        t3 = vec_mergel(src[4], src[5]);
+        t4 = vec_mergel(src[6], src[7]);
+
+        t5 = vec_xxpermdi(t1, t2, 0);
+        t6 = vec_xxpermdi(t3, t4, 0);
+        t7 = vec_xxpermdi(t1, t2, 3);
+        t8 = vec_xxpermdi(t3, t4, 3);
+
+        vec_xst(t5, 0, vecOffset + 16);
+        vec_xst(t6, 0, vecOffset + 20);
+        vec_xst(t7, 0, vecOffset + 24);
+        vec_xst(t8, 0, vecOffset + 28);
+    }
+
+    void packTranspose(const float * a, int64_t lda, int rows, int cols, float * vec) {
         int64_t i, j;
         float * aoffsets[8];
-        float *aoffset = NULL, *boffset = NULL;
+        float * aoffset = NULL, * boffset = NULL;
         __vector_pair arr[8];
         vector float c[8][2] = {0};
         vector float c1[8] = {0};
         vector float c2[8] = {0};
-        aoffset = const_cast<float*>(a);
+        aoffset = const_cast<float *>(a);
         boffset = vec;
         j = (rows >> 3);
         if (j > 0) {
-
             do {
                 aoffsets[0] = aoffset;
-                for (int it = 1; it< 8; it++)
+                for (int it = 1; it < 8; it++)
                     aoffsets[it] = aoffsets[it-1] + lda;
                 aoffset += 8 * lda;
                 i = (cols >> 3);
                 if (i > 0) {
                     do {
-                        for (int it = 0; it< 8; it++) {
+                        for (int it = 0; it < 8; it++) {
                             arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[it]);
                             __builtin_vsx_disassemble_pair(c[it], &arr[it]);
                             c1[it] = c[it][0];
@@ -2264,11 +2287,14 @@ class tinyBLAS_PPC {
                         }
 
                         vector_permute_store_8(c1, boffset);
-                        vector_permute_store_8(c2, boffset+32);
-                        for (int it = 0; it < 4; it++)
-                            aoffsets[it] = aoffsets[it] + 8*lda;
+                        vector_permute_store_8(c2, boffset + 32);
                         boffset += 64;
                         i--;
+                        if (i > 0) {
+                           for (int it = 0; it < 8; it++) {
+                               aoffsets[it] = aoffsets[it] + 8;
+                           }
+                        }
                     } while(i > 0);
                 }
                 if (cols & 4) {
@@ -2295,9 +2321,9 @@ class tinyBLAS_PPC {
                         c2[it] = c[it][1];
                     }
                     vector_permute_store_4(c1, boffset);
-                    vector_permute_store_4(c2, boffset+16);
+                    vector_permute_store_4(c2, boffset + 16);
                     for (int it = 0; it < 4; it++)
-                        aoffsets[it] += 8*lda;
+                        aoffsets[it] += 8 * lda;
                     boffset += 32;
                     i--;
                 } while(i > 0);
@@ -2325,15 +2351,15 @@ class tinyBLAS_PPC {
         vec_t vec_A[4], vec_B[4], vec_C[4];
         acc_t acc_0;
         __builtin_mma_xxsetaccz(&acc_0);
-        for (int l = 0; l < k; l+=4) {
-            packTranspose(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
-            packTranspose(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
+        for (int l = 0; l < k; l += 4) {
+            packTranspose(A + (ii * lda) + l, lda, 4, 4, (float *)vec_A);
+            packTranspose(B + (jj * ldb) + l, ldb, 4, 4, (float *)vec_B);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
         }
-        SAVE_ACC(&acc_0, ii, jj);
+        save_acc(&acc_0, ii, jj);
     }
 
     void KERNEL_4x8(int64_t ii, int64_t jj) {
@@ -2341,9 +2367,9 @@ class tinyBLAS_PPC {
         acc_t acc_0, acc_1;
         __builtin_mma_xxsetaccz(&acc_0);
         __builtin_mma_xxsetaccz(&acc_1);
-        for (int64_t l = 0; l < k; l+=4) {
-            packTranspose(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
-            packTranspose(B+(jj*ldb)+l, ldb, 8, 4, (float*)vec_B);
+        for (int64_t l = 0; l < k; l += 4) {
+            packTranspose(A + (ii * lda) + l, lda, 4, 4, (float *)vec_A);
+            packTranspose(B + (jj * ldb) + l, ldb, 8, 4, (float *)vec_B);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], (vec_t)vec_B[0]);
             __builtin_mma_xvf32gerpp(&acc_1, vec_A[0], (vec_t)vec_B[1]);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], (vec_t)vec_B[2]);
@@ -2353,8 +2379,8 @@ class tinyBLAS_PPC {
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[3], (vec_t)vec_B[6]);
             __builtin_mma_xvf32gerpp(&acc_1, vec_A[3], (vec_t)vec_B[7]);
         }
-        SAVE_ACC(&acc_0, ii, jj);
-        SAVE_ACC(&acc_1, ii, jj+4);
+        save_acc(&acc_0, ii, jj);
+        save_acc(&acc_1, ii, jj + 4);
     }
 
     void KERNEL_8x4(int64_t ii, int64_t jj) {
@@ -2362,9 +2388,9 @@ class tinyBLAS_PPC {
         acc_t acc_0, acc_1;
         __builtin_mma_xxsetaccz(&acc_0);
         __builtin_mma_xxsetaccz(&acc_1);
-        for (int64_t l = 0; l < k; l+=4) {
-            packTranspose(A+(ii*lda)+l, lda, 8, 4, (float*)vec_A);
-            packTranspose(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
+        for (int64_t l = 0; l < k; l += 4) {
+            packTranspose(A + (ii * lda) + l, lda, 8, 4, (float *)vec_A);
+            packTranspose(B + (jj * ldb) + l, ldb, 4, 4, (float *)vec_B);
             __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[0], vec_B[0]);
             __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[1], vec_B[0]);
             __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[2], vec_B[1]);
@@ -2374,8 +2400,8 @@ class tinyBLAS_PPC {
             __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[6], vec_B[3]);
             __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[7], vec_B[3]);
         }
-        SAVE_ACC(&acc_0, ii, jj);
-        SAVE_ACC(&acc_1, ii+4, jj);
+        save_acc(&acc_0, ii, jj);
+        save_acc(&acc_1, ii + 4, jj);
     }
 
     void KERNEL_8x8(int64_t ii, int64_t jj) {
@@ -2386,19 +2412,96 @@ class tinyBLAS_PPC {
         __builtin_mma_xxsetaccz(&acc_2);
         __builtin_mma_xxsetaccz(&acc_3);
         for (int l = 0; l < k; l+=8) {
-            packTranspose(A+(ii*lda)+l, lda, 8, 8, (float*)vec_A);
-            packTranspose(B+(jj*ldb)+l, ldb, 8, 8, (float*)vec_B);
+            packTranspose(A + (ii * lda) + l, lda, 8, 8, (float *)vec_A);
+            packTranspose(B + (jj * ldb) + l, ldb, 8, 8, (float *)vec_B);
             for(int x = 0; x < 16; x+=2) {
                 __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[x], vec_B[x]);
-                __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x+1]);
-                __builtin_mma_xvf32gerpp(&acc_2, (vec_t)vec_A[x+1], vec_B[x]);
-                __builtin_mma_xvf32gerpp(&acc_3, (vec_t)vec_A[x+1], vec_B[x+1]);
+                __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x + 1]);
+                __builtin_mma_xvf32gerpp(&acc_2, (vec_t)vec_A[x + 1], vec_B[x]);
+                __builtin_mma_xvf32gerpp(&acc_3, (vec_t)vec_A[x + 1], vec_B[x + 1]);
+            }
+        }
+        save_acc(&acc_0, ii, jj);
+        save_acc(&acc_1, ii, jj + 4);
+        save_acc(&acc_2, ii + 4, jj);
+        save_acc(&acc_3, ii + 4, jj + 4);
+    }
+
+    inline void MMA_16x8(vec_t * vec_A0, vec_t * vec_A1, vec_t * vec_B, acc_t * acc) {
+        for (int x = 0; x < 16; x += 2) {
+            __builtin_mma_xvf32gerpp(&acc[0], vec_A0[x + 0], vec_B[x]);
+            __builtin_mma_xvf32gerpp(&acc[1], vec_A0[x + 0], vec_B[x + 1]);
+            __builtin_mma_xvf32gerpp(&acc[2], vec_A0[x + 1], vec_B[x]);
+            __builtin_mma_xvf32gerpp(&acc[3], vec_A0[x + 1], vec_B[x + 1]);
+            __builtin_mma_xvf32gerpp(&acc[4], vec_A1[x + 0], vec_B[x]);
+            __builtin_mma_xvf32gerpp(&acc[5], vec_A1[x + 0], vec_B[x + 1]);
+            __builtin_mma_xvf32gerpp(&acc[6], vec_A1[x + 1], vec_B[x]);
+            __builtin_mma_xvf32gerpp(&acc[7], vec_A1[x + 1], vec_B[x + 1]);
+        }
+    }
+
+    void KERNEL(int64_t ii, int64_t jj, int64_t mc, int64_t nc, int64_t kc, vec_t * vec_A, vec_t * vec_B, int64_t kk) {
+        for (int64_t i = 0; i < mc; i += 16) {
+            int A_base_addr = (mc / 8) * (i / 8) * 16;
+            for (int64_t j = 0; j < nc; j += 8) {
+                 int B_base_addr = (nc / 8) * (j / 8) * 16;
+                 acc_t acc[8];
+                 vec_t A0_block[16]; vec_t A1_block[16];
+                 for (int x = 0; x < 8; x++)
+                     __builtin_mma_xxsetaccz(&acc[x]);
+                 for (int64_t l = 0; l < kc; l += 8) {
+                     int A0_block_idx = A_base_addr + (l / 8) * 16;
+                     int A1_block_idx = A0_block_idx + (mc / 8) * 16;
+                     int B_block_idx = B_base_addr + (l / 8) * 16;
+                     vec_t* A0_block = &vec_A[A0_block_idx];
+                     vec_t* A1_block = &vec_A[A1_block_idx];
+                     vec_t* B_block = &vec_B[B_block_idx];
+                     MMA_16x8(A0_block, A1_block, B_block, acc);
+                 }
+                 if (kk == 0) {
+                     save_acc(&acc[0], ii + i, jj + j);
+                     save_acc(&acc[1], ii + i, jj + j + 4);
+                     save_acc(&acc[2], ii + i + 4, jj + j);
+                     save_acc(&acc[3], ii + i + 4, jj + j + 4);
+                     save_acc(&acc[4], ii + i + 8, jj + j);
+                     save_acc(&acc[5], ii + i + 8, jj + j + 4);
+                     save_acc(&acc[6], ii + i + 12, jj + j);
+                     save_acc(&acc[7], ii + i + 12, jj + j + 4);
+                 } else {
+                     add_save_acc(&acc[0], ii + i, jj + j);
+                     add_save_acc(&acc[1], ii + i, jj + j + 4);
+                     add_save_acc(&acc[2], ii + i + 4, jj + j);
+                     add_save_acc(&acc[3], ii + i + 4, jj + j + 4);
+                     add_save_acc(&acc[4], ii + i + 8, jj + j);
+                     add_save_acc(&acc[5], ii + i + 8, jj + j + 4);
+                     add_save_acc(&acc[6], ii + i + 12, jj + j);
+                     add_save_acc(&acc[7], ii + i + 12, jj + j + 4);
+                 }
+            }
+        }
+    }
+
+    void matmul_tiled(int64_t m , int64_t n, int64_t mc, int64_t nc, int64_t kc) {
+        int64_t ytiles = m / mc;
+        int64_t xtiles = n / nc;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles) {
+            end = tiles;
+        }
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = (job / xtiles) * mc;
+            int64_t jj = (job % xtiles) * nc;
+            for (int64_t kk = 0; kk < k; kk += kc) {
+                 vec_t A_pack[kc * mc / 4];
+                 vec_t B_pack[kc * nc / 4];
+                 packTranspose(A + (ii * lda) + kk, lda, kc, mc, (float *)A_pack);
+                 packTranspose(B + (jj * ldb) + kk, ldb, kc, nc, (float *)B_pack);
+                 KERNEL(ii, jj, mc, nc, kc, A_pack, B_pack, kk);
             }
         }
-        SAVE_ACC(&acc_0, ii, jj);
-        SAVE_ACC(&acc_1, ii, jj+4);
-        SAVE_ACC(&acc_2, ii+4, jj);
-        SAVE_ACC(&acc_3, ii+4, jj+4);
     }
 
     void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
@@ -2406,35 +2509,35 @@ class tinyBLAS_PPC {
         int n_rem = MIN(n - n0, 8);
         int mc = 0, nc = 0;
         if (m_rem >= 8 && n_rem >= 8) {
-           mc = 8;
-           nc = 8;
-           gemm<8, 8>(m0, m, n0, n);
+            mc = 8;
+            nc = 8;
+            gemm<8, 8>(m0, m, n0, n);
         } else if (m_rem >= 4 && n_rem >= 8) {
-                mc = 4;
-                nc = 8;
-                gemm<4, 8>(m0, m, n0, n);
+            mc = 4;
+            nc = 8;
+            gemm<4, 8>(m0, m, n0, n);
         } else if (m_rem >= 8 && n_rem >= 4) {
-                mc = 8;
-                nc = 4;
-                gemm<8, 4>(m0, m, n0, n);
+            mc = 8;
+            nc = 4;
+            gemm<8, 4>(m0, m, n0, n);
         } else if (m_rem >= 4 && n_rem >= 4) {
-                mc = 4;
-                nc = 4;
-                gemm<4, 4>(m0, m, n0, n);
+            mc = 4;
+            nc = 4;
+            gemm<4, 4>(m0, m, n0, n);
         } else {
             mc = (m_rem >= 4) ? 4 : m_rem;
             nc = (n_rem >= 4) ? 4 : n_rem;
             if (mc == 0 || nc == 0)
-               return;
+                return;
             gemm_small(m0, m, n0, n, mc, nc);
         }
         int64_t mp = m0 + ((m - m0) / mc) * mc;
         int64_t np = n0 + ((n - n0) / nc) * nc;
         mnpack(mp, m, n0, np);
         mnpack(m0, m, np, n);
-     }
+    }
 
-     void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
+    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
         int64_t ytiles = (m - m0) / RM;
         int64_t xtiles = (n - n0) / RN;
         int64_t tiles = xtiles * ytiles;
@@ -2449,30 +2552,30 @@ class tinyBLAS_PPC {
             vec_t vec_C[4];
             acc_t acc_0;
             __builtin_mma_xxsetaccz(&acc_0);
-            vec_t vec_A[4] {0}, vec_B[4] = {0};
-            for (int l=0; l<k; l+=4) {
+            vec_t vec_A[4] = {0}, vec_B[4] = {0};
+            for (int l = 0; l < k; l += 4) {
                 /* 'GEMV Forwarding' concept is used in first two conditional loops.
                  * when one of the matrix has a single row/column, the elements are
                  * broadcasted, instead of using packing routine to prepack the
                  * matrix elements.
                  */
                 if (RM == 1) {
-                    float* a = const_cast<float*>(A+(ii)*lda+l);
-                    packTranspose(B+(jj*ldb)+l, ldb, RN, 4, (float*)vec_B);
+                    float * a = const_cast<float *>(A + (ii) * lda + l);
+                    packTranspose(B + (jj * ldb) + l, ldb, RN, 4, (float *)vec_B);
                     vec_A[0] = (vec_t)vec_xl(0,a);
-                    vec_A[1] = (vec_t)vec_splats(*((float*)&vec_A+1));
-                    vec_A[2] = (vec_t)vec_splats(*((float*)&vec_A+2));
-                    vec_A[3] = (vec_t)vec_splats(*((float*)&vec_A+3));
+                    vec_A[1] = (vec_t)vec_splats(*((float *)&vec_A+1));
+                    vec_A[2] = (vec_t)vec_splats(*((float *)&vec_A+2));
+                    vec_A[3] = (vec_t)vec_splats(*((float *)&vec_A+3));
                 } else if (RN == 1) {
-                    packTranspose(A+(ii*lda)+l, lda, RM, 4, (float*)vec_A);
-                    float* b = const_cast<float*>(B+(jj)*ldb+l);
+                    packTranspose(A + (ii * lda) + l, lda, RM, 4, (float *)vec_A);
+                    float * b = const_cast<float *>(B + (jj) * ldb + l);
                     vec_B[0] = (vec_t)vec_xl(0,b);
-                    vec_B[1] = (vec_t)vec_splats(*((float*)&vec_B+1));
-                    vec_B[2] = (vec_t)vec_splats(*((float*)&vec_B+2));
-                    vec_B[3] = (vec_t)vec_splats(*((float*)&vec_B+3));
+                    vec_B[1] = (vec_t)vec_splats(*((float *)&vec_B+1));
+                    vec_B[2] = (vec_t)vec_splats(*((float *)&vec_B+2));
+                    vec_B[3] = (vec_t)vec_splats(*((float *)&vec_B+3));
                 } else {
-                    packTranspose(A+(ii*lda)+l, lda, RM, 4, (float*)vec_A);
-                    packTranspose(B+(jj*ldb)+l, ldb, RN, 4, (float*)vec_B);
+                    packTranspose(A + (ii * lda) + l, lda, RM, 4, (float *)vec_A);
+                    packTranspose(B + (jj * ldb) + l, ldb, RN, 4, (float *)vec_B);
                 }
                 __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
                 __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
@@ -2482,12 +2585,27 @@ class tinyBLAS_PPC {
             __builtin_mma_disassemble_acc(vec_C, &acc_0);
             for (int I = 0; I < RM; I++) {
                 for (int J = 0; J < RN; J++) {
-                    *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J);
+                    *((float *)(C+ii+((jj+J)*ldc)+I)) = *((float *)&vec_C[I]+J);
                 }
             }
        }
     }
 
+    template<int RM, int RN>
+    inline void kernel(int64_t ii, int64_t jj) {
+        if constexpr(RM == 4 && RN == 4) {
+            KERNEL_4x4(ii, jj);
+        } else if constexpr(RM == 4 && RN == 8) {
+            KERNEL_4x8(ii, jj);
+        } else if constexpr(RM == 8 && RN == 4) {
+            KERNEL_8x4(ii, jj);
+        } else if constexpr(RM == 8 && RN == 8) {
+            KERNEL_8x8(ii, jj);
+        } else {
+            static_assert(false, "RN/RM values not supported");
+        }
+    }
+
     template <int RM, int RN>
     NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int64_t ytiles = (m - m0) / RM;
@@ -2496,27 +2614,18 @@ class tinyBLAS_PPC {
         int64_t duty = (tiles + nth - 1) / nth;
         int64_t start = duty * ith;
         int64_t end = start + duty;
-        if (RM == 4 && RN == 4) {
-            kernel = &tinyBLAS_PPC::KERNEL_4x4;
-        } else if (RM == 4 && RN == 8) {
-            kernel = &tinyBLAS_PPC::KERNEL_4x8;
-        } else if (RM == 8 && RN == 4) {
-            kernel = &tinyBLAS_PPC::KERNEL_8x4;
-        } else if (RM == 8 && RN == 8) {
-            kernel = &tinyBLAS_PPC::KERNEL_8x8;
-        }
         if (end > tiles)
             end = tiles;
         for (int64_t job = start; job < end; ++job) {
             int64_t ii = m0 + job / xtiles * RM;
             int64_t jj = n0 + job % xtiles * RN;
-            (this->*kernel)(ii, jj);
+            kernel<RM, RN>(ii, jj);
         }
     }
 
-    const float *const A;
-    const float *const B;
-    float *C;
+    const float * const A;
+    const float * const B;
+    float * C;
     const int64_t k;
     const int64_t lda;
     const int64_t ldb;

ggml/src/ggml-cpu/ops.cpp

@@ -7207,6 +7207,148 @@ void ggml_compute_forward_conv_2d(
     ggml_compute_forward_conv_2d_impl(params, src0, src1, dst, src0->type);
 }
 
+// ggml_compute_forward_conv_3d
+
+static void ggml_compute_forward_conv_3d_impl(const ggml_compute_params * params,
+                                              const ggml_tensor *         kernel,
+                                              const ggml_tensor *         src,
+                                              ggml_tensor *               dst,
+                                              ggml_type                   kernel_type) {
+
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(kernel_type == GGML_TYPE_F16 || kernel_type == GGML_TYPE_F32);
+    GGML_ASSERT(kernel->type == kernel_type);
+
+    const ggml_type_traits * traits = ggml_get_type_traits(kernel_type);
+
+    const int32_t s0 = dst->op_params[0];
+    const int32_t s1 = dst->op_params[1];
+    const int32_t s2 = dst->op_params[2];
+    const int32_t p0 = dst->op_params[3];
+    const int32_t p1 = dst->op_params[4];
+    const int32_t p2 = dst->op_params[5];
+    const int32_t d0 = dst->op_params[6];
+    const int32_t d1 = dst->op_params[7];
+    const int32_t d2 = dst->op_params[8];
+    const int32_t c  = dst->op_params[9];
+    const int32_t n  = dst->op_params[10];
+    const int32_t oc = dst->op_params[11];
+
+    const int64_t src_w = src->ne[0];
+    const int64_t src_h = src->ne[1];
+    const int64_t src_d = src->ne[2];
+    const int64_t knl_w = kernel->ne[0];
+    const int64_t knl_h = kernel->ne[1];
+    const int64_t knl_d = kernel->ne[2];
+    const int64_t dst_w = dst->ne[0];
+    const int64_t dst_h = dst->ne[1];
+    const int64_t dst_d = dst->ne[2];
+
+    const float * src_data = (float *) src->data;
+    void  * knl_data       = kernel->data;
+    float * dst_data       = (float *) dst->data;
+
+    const int64_t knl_n_per_channel = knl_w * knl_h * knl_d;
+    const int64_t knl_n_total       = knl_n_per_channel * c;
+    const int64_t patch_total       = n * dst_w * dst_h * dst_d;
+
+    const int64_t space_per_patch   = knl_n_total * traits->type_size + oc * sizeof(float);
+    const int64_t batch_size        = params->wsize / space_per_patch;
+    const int64_t patches_per_batch = batch_size > 8 ? (batch_size / 8) * 8 : batch_size;
+    const int64_t batch_n           = (patch_total + patches_per_batch - 1) / patches_per_batch;
+
+    GGML_ASSERT(patches_per_batch > 0 && batch_size >= 1);
+
+    void * tmp = params->wdata;
+
+    for (int64_t batch_i = 0; batch_i < batch_n; ++batch_i) {
+        const int64_t patch_start_batch = batch_i * patches_per_batch;
+        const int64_t patch_end_batch   = std::min(patch_start_batch + patches_per_batch, patch_total);
+        const int64_t patch_n_in_batch  = patch_end_batch - patch_start_batch;
+
+        const int64_t patch_per_thread  = (patch_n_in_batch + params->nth - 1) / params->nth;
+        const int64_t patch_start       = patch_start_batch + params->ith * patch_per_thread;
+        const int64_t patch_end         = std::min(patch_start + patch_per_thread, patch_end_batch);
+
+        for (int64_t p = patch_start; p < patch_end; ++p) {
+            const int64_t p_in_batch = p % (dst_w * dst_h * dst_d);
+            const int64_t p_in_depth = p_in_batch % (dst_w * dst_h);
+            const int64_t batch_idx  = p / (dst_w * dst_h * dst_d);
+            const int64_t dst_z      = p_in_batch / (dst_w * dst_h);
+            const int64_t dst_y      = p_in_depth / dst_w;
+            const int64_t dst_x      = p_in_depth % dst_w;
+
+            char * dst_row = (char *) tmp + (p % patches_per_batch) * knl_n_total * traits->type_size;
+
+            for (int64_t ic = 0; ic < c; ++ic) {
+                for (int64_t kz = 0; kz < knl_d; ++kz) {
+                    for (int64_t ky = 0; ky < knl_h; ++ky) {
+                        for (int64_t kx = 0; kx < knl_w; ++kx) {
+                            const int64_t sz = dst_z * s2 + kz * d2 - p2;
+                            const int64_t sy = dst_y * s1 + ky * d1 - p1;
+                            const int64_t sx = dst_x * s0 + kx * d0 - p0;
+
+                            int64_t dst_idx = ic * knl_n_per_channel + kz * (knl_h * knl_w) + ky * knl_w + kx;
+
+                            float src_val;
+                            if (sz < 0 || sz >= src_d || sy < 0 || sy >= src_h || sx < 0 || sx >= src_w) {
+                                src_val = 0.0f;
+                            } else {
+                                const int64_t cn_idx = batch_idx * c + ic;
+                                const float * src_ptr = (const float *)((const char *)src_data + sx*src->nb[0] + sy*src->nb[1] + sz*src->nb[2] + cn_idx*src->nb[3]);
+                                src_val = *src_ptr;
+                            }
+
+                            char * element_ptr = dst_row + dst_idx * traits->type_size;
+                            if (kernel_type == GGML_TYPE_F32) {
+                                *(float *)element_ptr = src_val;
+                            } else if (kernel_type == GGML_TYPE_F16) {
+                                *(ggml_fp16_t *)element_ptr = GGML_CPU_FP32_TO_FP16(src_val);
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        ggml_barrier(params->threadpool);
+
+        float * gemm_output = (float *) ((char *) tmp + patches_per_batch * knl_n_total * traits->type_size);
+        ggml_call_mul_mat(kernel_type, params, patch_n_in_batch, oc, knl_n_total, tmp, knl_data, gemm_output);
+
+        ggml_barrier(params->threadpool);
+
+        const int64_t permute_per_thread = (patch_n_in_batch + params->nth - 1) / params->nth;
+        const int64_t permute_start = params->ith * permute_per_thread;
+        const int64_t permute_end = std::min(permute_start + permute_per_thread, patch_n_in_batch);
+
+        for (int64_t i = permute_start; i < permute_end; ++i) {
+            const int64_t p = patch_start_batch + i;
+            const int64_t p_in_batch = p % (dst_w * dst_h * dst_d);
+            const int64_t p_in_depth = p_in_batch % (dst_w * dst_h);
+            const int64_t batch_idx  = p / (dst_w * dst_h * dst_d);
+            const int64_t dst_z      = p_in_batch / (dst_w * dst_h);
+            const int64_t dst_y      = p_in_depth / dst_w;
+            const int64_t dst_x      = p_in_depth % dst_w;
+
+            for (int64_t ioc = 0; ioc < oc; ++ioc) {
+                const float value = gemm_output[i * oc + ioc];
+                const int64_t ocn_idx = batch_idx * oc + ioc;
+                float * dst_ptr = (float *)((char *)dst_data + dst_x*dst->nb[0] + dst_y*dst->nb[1] + dst_z*dst->nb[2] + ocn_idx*dst->nb[3]);
+                *dst_ptr = value;
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_conv_3d(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    ggml_compute_forward_conv_3d_impl(params, src0, src1, dst, src0->type);
+}
+
 // ggml_compute_forward_conv_transpose_2d
 
 void ggml_compute_forward_conv_transpose_2d(
@@ -8930,6 +9072,9 @@ static void ggml_compute_forward_ssm_scan_f32(
                         }
 
                         sumf = GGML_F32xt_REDUCE_ONE(sum);
+    #elif defined(__riscv_v_intrinsic)
+                        // todo: RVV implementation
+                        const int np = 0;
     #else
                         const int np = (nc & ~(GGML_F32_STEP - 1));
 
@@ -9881,8 +10026,8 @@ static void ggml_compute_forward_rwkv_wkv7_f32(
     int64_t h_stride_2d = head_size * head_size;
 
     #if defined(GGML_SIMD)
-        #if defined(__ARM_FEATURE_SVE)
-            // scalar Route to scalar implementation       //TODO: Write SVE code
+        #if defined(__ARM_FEATURE_SVE) || defined(__riscv_v_intrinsic)
+            // scalar Route to scalar implementation       //TODO: Write SVE code and RVV code
             for (int64_t t = 0; t < T; t++) {
                 int64_t t_offset = t * t_stride;
                 int64_t state_offset = head_size * C * (t / (T / n_seqs));

ggml/src/ggml-cpu/ops.h

@@ -70,6 +70,7 @@ void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * p
 void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_conv_3d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);

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

@@ -18,6 +18,10 @@
 #include <immintrin.h>
 #endif
 
+#if defined(__riscv_v_intrinsic)
+#include <riscv_vector.h>
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -94,24 +98,15 @@ extern "C" {
     }
 #elif defined(__riscv) && defined(__riscv_zfhmin)
     static inline float riscv_compute_fp16_to_fp32(ggml_fp16_t h) {
-        float f;
-        __asm__(
-            "fmv.h.x %[f], %[h]\n\t"
-            "fcvt.s.h %[f], %[f]"
-            : [f] "=&f" (f)
-            : [h] "r" (h)
-        );
-        return f;
+        _Float16 hf;
+        memcpy(&hf, &h, sizeof(ggml_fp16_t));
+        return hf;
     }
 
     static inline ggml_fp16_t riscv_compute_fp32_to_fp16(float f) {
         ggml_fp16_t res;
-        __asm__(
-            "fcvt.h.s %[f], %[f]\n\t"
-            "fmv.x.h %[h], %[f]"
-            : [h] "=&r" (res)
-            : [f] "f" (f)
-        );
+        _Float16 hf = (_Float16)f;
+        memcpy(&res, &hf, sizeof(ggml_fp16_t));
         return res;
     }
 
@@ -1170,6 +1165,36 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
 #define GGML_F16_VEC_MUL            GGML_F32x4_MUL
 #define GGML_F16_VEC_REDUCE         GGML_F32x4_REDUCE
 
+#elif defined(__riscv_v_intrinsic)
+
+// compatible with vlen >= 128
+
+#define GGML_SIMD
+
+// F32
+
+#define GGML_F32_STEP 16
+#define GGML_F32_EPR  4
+
+#define GGML_F32x4              vfloat32m1_t
+#define GGML_F32x4_ZERO         __riscv_vfmv_v_f_f32m1(0.0f, GGML_F32_EPR)
+#define GGML_F32x4_SET1(x)      __riscv_vfmv_v_f_f32m1(x, GGML_F32_EPR)
+#define GGML_F32x4_LOAD(x)      __riscv_vle32_v_f32m1(x, GGML_F32_EPR)
+#define GGML_F32x4_STORE(b, v)  __riscv_vse32_v_f32m1(b, v, GGML_F32_EPR)
+#define GGML_F32x4_FMA(a, b, c) __riscv_vfmacc_vv_f32m1(a, b, c, GGML_F32_EPR)
+#define GGML_F32x4_ADD(a, b)    __riscv_vfadd_vv_f32m1(a, b, GGML_F32_EPR)
+#define GGML_F32x4_MUL(a, b)    __riscv_vfmul_vv_f32m1(a, b, GGML_F32_EPR)
+
+#define GGML_F32_VEC        GGML_F32x4
+#define GGML_F32_VEC_ZERO   GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA    GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD    GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL    GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
 #endif
 
 // GGML_F32_ARR / GGML_F16_ARR

ggml/src/ggml-cpu/vec.cpp

@@ -84,6 +84,16 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
         }
         // reduce sum1,sum2 to sum1
         GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
+    #elif defined(__riscv_v_intrinsic)
+        vfloat32m1_t vsum = __riscv_vfmv_v_f_f32m1(0.0f, 1);
+        for (int i = 0, avl; i < n; i += avl) {
+            avl = __riscv_vsetvl_e32m8(n - i);
+            vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
+            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+            vfloat32m8_t prod = __riscv_vfmul_vv_f32m8(ax, ay, avl);
+            vsum = __riscv_vfredusum_vs_f32m8_f32m1(prod, vsum, avl);
+        }
+        sumf += __riscv_vfmv_f_s_f32m1_f32(vsum);
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -197,7 +207,7 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
 
     ggml_float sumf = 0.0;
 
-#if defined(GGML_SIMD)
+#if defined(GGML_SIMD) && !defined(__riscv_v_intrinsic)
     const int np = (n & ~(GGML_F16_STEP - 1));
 
     GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO };
@@ -325,6 +335,15 @@ ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float
         vst1q_f32(y + i, val);
         sum += (ggml_float)vaddvq_f32(val);
     }
+#elif defined(__riscv_v_intrinsic)
+    vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
+    for (int avl; i < n; i += avl) {
+        avl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t val = ggml_v_expf_m2(__riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], avl), max, avl), avl);
+        __riscv_vse32_v_f32m2(&y[i], val, avl);
+        vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, avl);
+    }
+    return (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
 #endif
     for (; i < n; ++i) {
         float val = expf(x[i] - max);

ggml/src/ggml-cpu/vec.h

@@ -119,6 +119,14 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
     }
 
 #if defined(GGML_SIMD)
+#if defined(__riscv_v_intrinsic)
+    // todo: RVV impl
+    for (int i = 0; i < n; ++i) {
+        for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
+            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
+        }
+    }
+#else
     const int np = (n & ~(GGML_F16_STEP - 1));
 
     GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } };
@@ -149,6 +157,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
             sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
         }
     }
+#endif
 #else
     for (int i = 0; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
@@ -243,6 +252,14 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 
             svst1_f32(pg, y + np2, ay1);
         }
+    #elif defined(__riscv_v_intrinsic)
+        for (int i = 0, avl; i < n; i += avl) {
+            avl = __riscv_vsetvl_e32m8(n - i);
+            vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
+            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+            vfloat32m8_t ny = __riscv_vfmadd_vf_f32m8(ax, v, ay, avl);
+            __riscv_vse32_v_f32m8(&y[i], ny, avl);
+        }
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -276,6 +293,13 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 
 inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD)
+#if defined(__riscv_v_intrinsic)
+    // todo: RVV impl
+    // scalar
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
+    }
+#else
     const int np = (n & ~(GGML_F16_STEP - 1));
 
     GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
@@ -297,6 +321,7 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
     for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
     }
+#endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -324,6 +349,16 @@ inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int
                 y[i] += x[k][i]*v[k][0];
             }
         }
+    #elif defined(__riscv_v_intrinsic)
+        for (int i = 0, avl; i < n; i += avl) {
+            avl = __riscv_vsetvl_e32m8(n - i);
+            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+            for (int k = 0; k < GGML_VEC_MAD_UNROLL; k++) {
+                vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[k][i], avl);
+                ay = __riscv_vfmadd_vf_f32m8(ax, v[k][0], ay, avl);
+            }
+            __riscv_vse32_v_f32m8(&y[i], ay, avl);
+        }
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -375,6 +410,14 @@ inline static void ggml_vec_mad1_f32(const int n, float * y, const float * x, co
         for (int i = 0; i < n; ++i) {
             y[i] = x[i]*s + b;
         }
+    #elif defined(__riscv_v_intrinsic)
+        for (int i = 0, avl; i < n; i += avl) {
+            avl = __riscv_vsetvl_e32m8(n - i);
+            vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
+            vfloat32m8_t vb = __riscv_vfmv_v_f_f32m8(b, avl);
+            vfloat32m8_t ny = __riscv_vfmadd_vf_f32m8(ax, s, vb, avl);
+            __riscv_vse32_v_f32m8(&y[i], ny, avl);
+        }
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -436,6 +479,13 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
             ay1 = svmul_f32_m(pg, ay1, vx);
             svst1_f32(pg, y + np, ay1);
         }
+    #elif defined(__riscv_v_intrinsic)
+        for (int i = 0, avl; i < n; i += avl) {
+            avl = __riscv_vsetvl_e32m8(n - i);
+            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+            vfloat32m8_t ny = __riscv_vfmul_vf_f32m8(ay, v, avl);
+            __riscv_vse32_v_f32m8(&y[i], ny, avl);
+        }
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -467,6 +517,13 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 
 inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
 #if defined(GGML_SIMD)
+#if defined(__riscv_v_intrinsic)
+    // todo: RVV impl
+    // scalar
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
+    }
+#else
     const int np = (n & ~(GGML_F16_STEP - 1));
 
     GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
@@ -486,6 +543,7 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
     for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
     }
+#endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -928,7 +986,51 @@ inline static __m128 ggml_v_silu(__m128 x) {
     return _mm_div_ps(x, one_plus_exp_neg_x);
 }
 
-#endif // __ARM_NEON / __AVX2__ / __SSE2__
+#elif defined(__riscv_v_intrinsic)
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static vfloat32m2_t ggml_v_expf_m2(vfloat32m2_t x, int vl) {
+    const vfloat32m2_t r = __riscv_vfmv_v_f_f32m2(0x1.8p23f, vl);
+#ifdef __riscv_xtheadvector
+    // workaround for compiler bug (gcc 14.3.0: Error: unrecognized opcode `th.vmv1r.v v2,v4')
+    vfloat32m2_t z = __riscv_vfadd_vf_f32m2(r, 0.0f, vl);
+    z = __riscv_vfmacc_vf_f32m2(z, 0x1.715476p+0f, x, vl);
+#else
+    const vfloat32m2_t z = __riscv_vfmacc_vf_f32m2(r, 0x1.715476p+0f, x, vl);
+#endif
+    const vfloat32m2_t n = __riscv_vfsub_vv_f32m2(z, r, vl);
+    const vfloat32m2_t b = __riscv_vfnmsac_vf_f32m2(__riscv_vfnmsac_vf_f32m2(x, 0x1.62e4p-1f, n, vl),
+                                                    0x1.7f7d1cp-20f, n, vl);
+    const vuint32m2_t e = __riscv_vsll_vx_u32m2(__riscv_vreinterpret_v_f32m2_u32m2(z), 23, vl);
+    const vfloat32m2_t k = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vadd_vx_u32m2(e, 0x3f800000, vl)); // 1.0f
+    const vbool16_t c = __riscv_vmfgt_vf_f32m2_b16(__riscv_vfabs_v_f32m2(n, vl), 126.0f, vl);
+    const vfloat32m2_t u = __riscv_vfmul_vv_f32m2(b, b, vl);
+    const vfloat32m2_t j = __riscv_vfmacc_vv_f32m2(
+        __riscv_vfmul_vf_f32m2(b, 0x1.ffffecp-1f, vl),
+        __riscv_vfmacc_vv_f32m2(
+            __riscv_vfmacc_vf_f32m2(__riscv_vfmv_v_f_f32m2(0x1.fffdb6p-2f, vl), 0x1.555e66p-3f, b, vl),
+            __riscv_vfmacc_vf_f32m2(__riscv_vfmv_v_f_f32m2(0x1.573e2ep-5f, vl), 0x1.0e4020p-7f, b, vl),
+            u, vl), u, vl);
+    if (!__riscv_vcpop_m_b16(c, vl))
+        return __riscv_vfmacc_vv_f32m2(k, j, k, vl);
+    const vbool16_t  dm = __riscv_vmfle_vf_f32m2_b16(n, 0.0f, vl);
+    const vuint32m2_t d = __riscv_vmerge_vxm_u32m2(__riscv_vmv_v_x_u32m2(0, vl), 0x82000000, dm, vl);
+    const vfloat32m2_t s1 = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vadd_vx_u32m2(d, 0x7f000000, vl));
+    const vfloat32m2_t s2 = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vsub_vv_u32m2(e, d, vl));
+    const vfloat32m2_t r1 = __riscv_vmerge_vvm_f32m2(
+        __riscv_vfmacc_vv_f32m2(k, k, j, vl),
+        __riscv_vfmul_vv_f32m2(__riscv_vfmacc_vv_f32m2(s2, s2, j, vl), s1, vl),
+        c, vl);
+    return __riscv_vmerge_vvm_f32m2(
+        r1, __riscv_vfmul_vv_f32m2(s1, s1, vl),
+        __riscv_vmfgt_vf_f32m2_b16(__riscv_vfabs_v_f32m2(n, vl), 192.0f, vl),
+        vl);
+}
+
+#endif // __ARM_NEON / __AVX2__ / __SSE2__ / __riscv_v_intrinsic
 
 inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {

ggml/src/ggml-cuda/common.cuh

@@ -107,9 +107,9 @@ constexpr bool ggml_cuda_has_arch(const int arch) {
     return ggml_cuda_has_arch_impl(arch, __CUDA_ARCH_LIST__);
 }
 
-constexpr int ggml_cuda_highest_compiled_arch_impl(const int arch, const int cur) {
+constexpr int ggml_cuda_highest_compiled_arch_impl(const int /*arch*/, const int cur) {
     if (cur == 0) {
-        GGML_ABORT("ggml was not compiled with any CUDA arch <= %d", arch);
+        return -1;
     }
     return cur;
 }
@@ -420,16 +420,28 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_all(int x) {
-#ifdef GGML_USE_HIP
+    if (width == ggml_cuda_get_physical_warp_size()) {
+        return __all_sync(0xffffffff, x);
+    } else {
 #pragma unroll
-    for (int offset = width/2; offset > 0; offset >>= 1) {
-        x = x && __shfl_xor_sync(0xffffffff, x, offset, width);
+        for (int offset = width/2; offset > 0; offset >>= 1) {
+            x = __shfl_xor_sync(0xffffffff, x, offset, width) && x;
+        }
+        return x;
+    }
+}
+
+template<int width = WARP_SIZE>
+static __device__ __forceinline__ int warp_reduce_any(int x) {
+    if (width == ggml_cuda_get_physical_warp_size()) {
+        return __any_sync(0xffffffff, x);
+    } else {
+#pragma unroll
+        for (int offset = width/2; offset > 0; offset >>= 1) {
+            x = __shfl_xor_sync(0xffffffff, x, offset, width) || x;
+        }
+        return x;
     }
-    return x;
-#else
-    static_assert(width == WARP_SIZE, "width != WARP_SIZE not implemented");
-    return __all_sync(0xffffffff, x);
-#endif // GGML_USE_HIP
 }
 
 template<int width = WARP_SIZE>

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

@@ -258,7 +258,7 @@ static __global__ void flash_attn_tile_ext_f16(
             const half val = hexp(sink - kqmax[j0/nwarps]);
             kqsum[j0/nwarps] = kqsum[j0/nwarps] * KQ_max_scale;
             if (threadIdx.x == 0) {
-                kqsum[j0/nwarps].x = __hadd(kqsum[j0/nwarps].x, val);
+                kqsum[j0/nwarps].x = __hadd(__low2half(kqsum[j0/nwarps]), val);
             }
 
 #pragma unroll

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

@@ -49,6 +49,7 @@
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
 #include "ggml-cuda/set-rows.cuh"
+#include "ggml-cuda/pad_reflect_1d.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -203,6 +204,8 @@ static ggml_cuda_device_info ggml_cuda_init() {
     GGML_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: no\n", __func__);
 #endif // GGML_CUDA_FORCE_CUBLAS
     GGML_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
+
+    std::vector<std::pair<int, std::string>> turing_devices_without_mma;
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
@@ -260,7 +263,25 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
         GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
                         id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
-#endif // defined(GGML_USE_HIP)
+        std::string device_name(prop.name);
+        if (device_name == "NVIDIA GeForce MX450") {
+            turing_devices_without_mma.push_back({ id, device_name });
+        } else if (device_name == "NVIDIA GeForce MX550") {
+            turing_devices_without_mma.push_back({ id, device_name });
+        } else if (device_name.substr(0, 21) == "NVIDIA GeForce GTX 16") {
+            turing_devices_without_mma.push_back({ id, device_name });
+        }
+#endif  // defined(GGML_USE_HIP)
+    }
+
+    if (ggml_cuda_highest_compiled_arch(GGML_CUDA_CC_TURING) >= GGML_CUDA_CC_TURING && !turing_devices_without_mma.empty()) {
+        GGML_LOG_INFO("The following devices will have suboptimal performance due to a lack of tensor cores:\n");
+        for (size_t device_pos = 0; device_pos < turing_devices_without_mma.size(); device_pos++) {
+            GGML_LOG_INFO(
+                "  Device %d: %s\n", turing_devices_without_mma[device_pos].first, turing_devices_without_mma[device_pos].second.c_str());
+        }
+        GGML_LOG_INFO(
+            "Consider compiling with CMAKE_CUDA_ARCHITECTURES=61-virtual;80-virtual and DGGML_CUDA_FORCE_MMQ to force the use of the Pascal code for Turing.\n");
     }
 
     for (int id = 0; id < info.device_count; ++id) {
@@ -2352,6 +2373,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_PAD:
             ggml_cuda_op_pad(ctx, dst);
             break;
+        case GGML_OP_PAD_REFLECT_1D:
+            ggml_cuda_op_pad_reflect_1d(ctx, dst);
+            break;
         case GGML_OP_ARANGE:
             ggml_cuda_op_arange(ctx, dst);
             break;
@@ -3481,15 +3505,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
-        case GGML_OP_SUM_ROWS:
-        case GGML_OP_MEAN:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
             return true;
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:

ggml/src/ggml-cuda/mmq.cu

@@ -3,6 +3,140 @@
 
 #include <vector>
 
+// To reduce shared memory use, store "it" and "iex_used" with 22/10 bits each.
+struct mmq_ids_helper_store {
+    uint32_t data;
+
+    __device__ mmq_ids_helper_store(const uint32_t it, const uint32_t iex_used) {
+        data = (it & 0x003FFFFF) | (iex_used << 22);
+    }
+
+    __device__ uint32_t it() const {
+        return data & 0x003FFFFF;
+    }
+
+    __device__ uint32_t iex_used() const {
+        return data >> 22;
+    }
+};
+static_assert(sizeof(mmq_ids_helper_store) == 4, "unexpected size for mmq_ids_helper_store");
+
+// Helper function for mul_mat_id, converts ids to a more convenient format.
+// ids_src1 describes how to permute the flattened column indices of src1 in order to get a compact src1 tensor sorted by expert.
+// ids_dst describes the same mapping but for the dst tensor.
+// The upper and lower bounds for the ith expert in the compact src1 tensor are stored in expert_bounds[i:i+1].
+template <int n_expert_used_template>
+__launch_bounds__(ggml_cuda_get_physical_warp_size(), 1)
+static __global__ void mmq_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1) {
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
+    const int expert = blockIdx.x;
+
+    extern __shared__ char data_mmq_ids_helper[];
+    mmq_ids_helper_store * store = (mmq_ids_helper_store *) data_mmq_ids_helper;
+
+    int nex_prev   = 0; // Number of columns for experts with a lower index.
+    int it_compact = 0; // Running index for the compact slice of this expert.
+
+    if constexpr (n_expert_used_template == 0) {
+        // Generic implementation:
+        for (int it = 0; it < n_tokens; ++it) {
+            int iex_used = -1; // The index at which the expert is used, if any.
+            for (int iex = threadIdx.x; iex < n_expert_used; iex += warp_size) {
+                const int expert_used = ids[it*si1 + iex];
+                nex_prev += expert_used < expert;
+                if (expert_used == expert) {
+                    iex_used = iex;
+                }
+            }
+
+            if (iex_used != -1) {
+                store[it_compact] = mmq_ids_helper_store(it, iex_used);
+            }
+
+            if (warp_reduce_any<warp_size>(iex_used != -1)) {
+                it_compact++;
+            }
+        }
+    } else {
+        // Implementation optimized for specific numbers of experts used:
+        static_assert(n_expert_used == 6 || warp_size % n_expert_used == 0, "bad n_expert_used");
+        const int neu_padded = n_expert_used == 6 ? 8 : n_expert_used; // Padded to next higher power of 2.
+        for (int it0 = 0; it0 < n_tokens; it0 += warp_size/neu_padded) {
+            const int it = it0 + threadIdx.x / neu_padded;
+
+            const int iex = threadIdx.x % neu_padded; // The index at which the expert is used, if any.
+            const int expert_used = (neu_padded == n_expert_used || iex < n_expert_used) && it < n_tokens ?
+                ids[it*si1 + iex] : INT_MAX;
+            const int iex_used = expert_used == expert ? iex : -1;
+            nex_prev += expert_used < expert;
+
+            // Whether the threads at this token position have used the expert:
+            const int it_compact_add_self = warp_reduce_any<neu_padded>(iex_used != -1);
+
+            // Do a scan over threads at lower token positions in warp to get the correct index for writing data:
+            int it_compact_add_lower = 0;
+#pragma unroll
+            for (int offset = neu_padded; offset < warp_size; offset += neu_padded) {
+                const int tmp = __shfl_up_sync(0xFFFFFFFF, it_compact_add_self, offset, warp_size);
+                if (threadIdx.x >= offset) {
+                    it_compact_add_lower += tmp;
+                }
+            }
+
+            if (iex_used != -1) {
+                store[it_compact + it_compact_add_lower] = mmq_ids_helper_store(it, iex_used);
+            }
+
+            // The thread with the highest index in the warp always has the sum over the whole warp, use it to increment all threads:
+            it_compact += __shfl_sync(0xFFFFFFFF, it_compact_add_lower + it_compact_add_self, warp_size - 1, warp_size);
+        }
+    }
+    nex_prev = warp_reduce_sum<warp_size>(nex_prev);
+
+    for (int itc = threadIdx.x; itc < it_compact; itc += warp_size) {
+        const mmq_ids_helper_store store_it = store[itc];
+        const int it       = store_it.it();
+        const int iex_used = store_it.iex_used();
+        ids_src1[nex_prev + itc] = it*sis1          + iex_used % nchannels_y;
+        ids_dst [nex_prev + itc] = it*n_expert_used + iex_used;
+    }
+
+    if (threadIdx.x != 0) {
+        return;
+    }
+
+    expert_bounds[expert] = nex_prev;
+
+    if (expert < gridDim.x - 1) {
+        return;
+    }
+
+    expert_bounds[gridDim.x] = nex_prev + it_compact;
+}
+
+template <int n_expert_used_template>
+static void launch_mmq_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
+    GGML_ASSERT(n_tokens          < (1 << 22) && "too few bits in mmq_ids_helper_store");
+    GGML_ASSERT(n_expert_used_var < (1 << 10) && "too few bits in mmq_ids_helper_store");
+
+    const int id = ggml_cuda_get_device();
+    const int warp_size = ggml_cuda_info().devices[id].warp_size;
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+    CUDA_SET_SHARED_MEMORY_LIMIT(mmq_ids_helper<n_expert_used_template>, smpbo);
+
+    const dim3 num_blocks(n_experts, 1, 1);
+    const dim3 block_size(warp_size, 1, 1);
+    const size_t nbytes_shared = n_tokens*sizeof(mmq_ids_helper_store);
+    GGML_ASSERT(nbytes_shared <= smpbo);
+    mmq_ids_helper<n_expert_used_template><<<num_blocks, block_size, nbytes_shared, stream>>>
+        (ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1);
+}
+
 static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
     switch (args.type_x) {
         case GGML_TYPE_Q4_0:
@@ -137,7 +271,7 @@ void ggml_cuda_mul_mat_q(
             ne00, ne01, ne1, s01, ne11, s1,
             ne02, ne12, s02, s12, s2,
             ne03, ne13, s03, s13, s3,
-            use_stream_k};
+            use_stream_k, ne1};
         ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
         return;
     }
@@ -148,54 +282,50 @@ void ggml_cuda_mul_mat_q(
 
     const int64_t n_expert_used = ids->ne[0];
     const int64_t ne_get_rows = ne12 * n_expert_used;
+    GGML_ASSERT(ne1 == n_expert_used);
 
-    std::vector<char> ids_host(ggml_nbytes(ids));
-    std::vector<int32_t> ids_src1_host;
-    ids_src1_host.reserve(ne_get_rows);
-    std::vector<int32_t> ids_dst_host;
-    ids_dst_host.reserve(ne_get_rows);
-    std::vector<int32_t> tokens_per_expert_host(ne02);
-    std::vector<int32_t> expert_bounds_host(ne02 + 1);
-    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool());
+    ggml_cuda_pool_alloc<int32_t> ids_src1(ctx.pool(), ne_get_rows);
+    ggml_cuda_pool_alloc<int32_t> ids_dst(ctx.pool(), ne_get_rows);
+    ggml_cuda_pool_alloc<int32_t> expert_bounds(ctx.pool(), ne02 + 1);
 
-    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
-    CUDA_CHECK(cudaStreamSynchronize(stream));
+    {
+        GGML_ASSERT(ids->nb[0] == ggml_element_size(ids));
+        const int si1  = ids->nb[1] / ggml_element_size(ids);
+        const int sis1 = nb12 / nb11;
 
-    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
-        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
-            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
-                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
-                assert(expert_to_use >= 0 && expert_to_use < ne02);
-                if (expert_to_use == i02) {
-                    ids_src1_host.push_back(i12*(nb12/nb11) + iex % ne11);
-                    ids_dst_host.push_back(i12*ne1 + iex);
-                    tokens_per_expert_host[i02]++;
-                    break;
-                }
-            }
+        switch (n_expert_used) {
+            case  2:
+                launch_mmq_ids_helper< 2> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            case  4:
+                launch_mmq_ids_helper< 4> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            case  6:
+                launch_mmq_ids_helper< 6> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            case  8:
+                launch_mmq_ids_helper< 8> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            case 16:
+                launch_mmq_ids_helper<16> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            case 32:
+                launch_mmq_ids_helper<32> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            default:
+                launch_mmq_ids_helper< 0> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
         }
+        CUDA_CHECK(cudaGetLastError());
     }
 
-    int32_t cumsum = 0;
-    for (int64_t i = 0; i < ne02; ++i) {
-        expert_bounds_host[i] = cumsum;
-        cumsum += tokens_per_expert_host[i];
-    }
-    expert_bounds_host[ne02] = cumsum;
-
-    std::vector<int32_t> ids_buf_host;
-    ids_buf_host.reserve(ids_src1_host.size() + ids_dst_host.size() + expert_bounds_host.size());
-    ids_buf_host.insert(ids_buf_host.end(), ids_src1_host.begin(), ids_src1_host.end());
-    ids_buf_host.insert(ids_buf_host.end(), ids_dst_host.begin(), ids_dst_host.end());
-    ids_buf_host.insert(ids_buf_host.end(), expert_bounds_host.begin(), expert_bounds_host.end());
-    ids_buf_dev.alloc(ids_buf_host.size() + get_mmq_x_max_host(cc)); // Expert bounds are padded on device.
-    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_buf_host.data(), ids_buf_host.size()*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
-    CUDA_CHECK(cudaStreamSynchronize(stream));
-
-    const int32_t * ids_src1_dev      = ids_buf_dev.ptr;
-    const int32_t * ids_dst_dev       = ids_src1_dev + ids_src1_host.size();
-    const int32_t * expert_bounds_dev = ids_dst_dev + ids_dst_host.size();
-
     const size_t nbytes_src1_q8_1 = ne12*n_expert_used*ne10_padded * sizeof(block_q8_1)/QK8_1 +
         get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
     ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
@@ -208,7 +338,7 @@ void ggml_cuda_mul_mat_q(
         const int64_t s11 = src1->nb[1] / ts_src1;
         const int64_t s12 = src1->nb[2] / ts_src1;
         const int64_t s13 = src1->nb[2] / ts_src1;
-        quantize_mmq_q8_1_cuda(src1_d, ids_src1_dev, src1_q8_1.get(), src0->type,
+        quantize_mmq_q8_1_cuda(src1_d, ids_src1.get(), src1_q8_1.get(), src0->type,
             ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
         CUDA_CHECK(cudaGetLastError());
     }
@@ -218,11 +348,11 @@ void ggml_cuda_mul_mat_q(
 
     // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
     const mmq_args args = {
-        src0_d, src0->type, (const int *) src1_q8_1.ptr, ids_dst_dev, expert_bounds_dev, dst_d,
+        src0_d, src0->type, (const int *) src1_q8_1.get(), ids_dst.get(), expert_bounds.get(), dst_d,
         ne00, ne01, ne_get_rows, s01, ne_get_rows, s1,
         ne02, ne02, s02, s12, s2,
         ne03, ne13, s03, s13, s3,
-        use_stream_k};
+        use_stream_k, ne12};
 
     ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 }
@@ -262,7 +392,7 @@ void ggml_cuda_op_mul_mat_q(
         ne00, row_diff, src1_ncols, stride01, ne11, nrows_dst,
         1, 1, 0, 0, 0,
         1, 1, 0, 0, 0,
-        use_stream_k};
+        use_stream_k, src1_ncols};
 
     ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 

ggml/src/ggml-cuda/mmq.cuh

@@ -3138,7 +3138,8 @@ static __global__ void mul_mat_q(
         const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup,
         const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_row_x, const int ncols_y, const int stride_col_dst,
         const int channel_ratio, const int nchannels_y, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
-        const int sample_ratio, const int nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+        const int sample_ratio, const int nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const int ncols_max) {
 
     // Skip unused template specializations for faster compilation:
     if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) {
@@ -3152,7 +3153,7 @@ static __global__ void mul_mat_q(
     constexpr int qk    = ggml_cuda_type_traits<type>::qk;
     constexpr int mmq_y = get_mmq_y_device();
 
-    const int ntx = (ncols_dst + mmq_x - 1) / mmq_x; // Number of tiles x
+    const int ntx = (ncols_max + mmq_x - 1) / mmq_x; // Number of tiles x
     const int nty = (nrows_x   + mmq_y - 1) / mmq_y; // Number of tiles y
 
     // Initialize the ids for writing back data with just the index.
@@ -3376,7 +3377,8 @@ template <ggml_type type, int mmq_x, bool need_check>
 static __global__ void mul_mat_q_stream_k_fixup(
         const int32_t * ids_dst, const int32_t * expert_bounds, float * __restrict__ dst, const float * __restrict__ tmp_last_tile,
         const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_col_dst,
-        const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst) {
+        const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst,
+        const int ncols_max) {
     constexpr int     mmq_y           = get_mmq_y_device();
     constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
@@ -3387,7 +3389,7 @@ static __global__ void mul_mat_q_stream_k_fixup(
 
     float sum[mmq_x*mmq_y / (nwarps*warp_size)] = {0.0f};
 
-    const int ntx  = (ncols_dst + mmq_x - 1) / mmq_x;
+    const int ntx  = (ncols_max + mmq_x - 1) / mmq_x;
     const int nty  = (nrows_x   + mmq_y - 1) / mmq_y;
 
     const int bidx0 = blockIdx.x;
@@ -3528,7 +3530,7 @@ struct mmq_args {
     int64_t ncols_x; int64_t nrows_x; int64_t ncols_dst; int64_t stride_row_x; int64_t ncols_y; int64_t nrows_dst;
     int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst;
     int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst;
-    bool use_stream_k;
+    bool use_stream_k; int64_t ncols_max;
 };
 
 template<ggml_type type>
@@ -3558,7 +3560,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
     CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x,  true>), nbytes_shared);
 
     const int nty  = (args.nrows_x   + mmq_y - 1) / mmq_y;
-    const int ntx  = (args.ncols_dst + mmq_x - 1) / mmq_x;
+    const int ntx  = (args.ncols_max + mmq_x - 1) / mmq_x;
     const int ntzw = args.nchannels_y * args.nsamples_y;
     const dim3 block_nums_xy_tiling(nty, ntx, ntzw);
 
@@ -3574,14 +3576,16 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
                 (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
                  args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
                  channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
+                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+                 args.ncols_max);
         } else {
             constexpr bool need_check = true;
             mul_mat_q<type, mmq_x, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
                 (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
                  args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
                  channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
+                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+                 args.ncols_max);
         }
         return;
     }
@@ -3601,7 +3605,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
             (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
              args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
              channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
+             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+             args.ncols_max);
 
         if (!fixup_needed) {
             return;
@@ -3609,14 +3614,16 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
         mul_mat_q_stream_k_fixup<type, mmq_x, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
             (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_dst,
-             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
+             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst,
+             args.ncols_max);
     } else {
         constexpr bool need_check = true;
         mul_mat_q<type, mmq_x, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
             (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
              args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
              channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
+             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+             args.ncols_max);
 
         if (!fixup_needed) {
             return;
@@ -3624,7 +3631,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
         mul_mat_q_stream_k_fixup<type, mmq_x, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
             (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_dst,
-             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
+             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst,
+             args.ncols_max);
     }
 }
 
@@ -3649,7 +3657,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
             continue;
         }
 
-        const int ntiles_x = (args.ncols_y + mmq_x - 1) / mmq_x;
+        const int ntiles_x = (args.ncols_max + mmq_x - 1) / mmq_x;
 
         if (ntiles_x < ntiles_x_best) {
             mmq_x_best = mmq_x;

ggml/src/ggml-cuda/pad_reflect_1d.cu

@@ -0,0 +1,82 @@
+#include "pad_reflect_1d.cuh"
+
+static __global__ void pad_reflect_1d_kernel_f32(
+    const void * __restrict__ src0,
+    void * __restrict__ dst,
+    const int64_t ne0,
+    const int64_t ne00,
+    const int64_t ne01,
+    const int64_t ne02,
+    const int64_t ne03,
+    const int64_t nb00,
+    const int64_t nb01,
+    const int64_t nb02,
+    const int64_t nb03,
+    const int64_t nb0,
+    const int64_t nb1,
+    const int64_t nb2,
+    const int64_t nb3,
+    const int p0,
+    const int p1) {
+
+    const int64_t i3 = blockIdx.z;
+    const int64_t i2 = blockIdx.y;
+    const int64_t i1 = blockIdx.x;
+
+    if (i1 >= ne01 || i2 >= ne02 || i3 >= ne03) {
+        return;
+    }
+
+    const char * src0_ptr = (const char *)src0 + i3*nb03 + i2*nb02 + i1*nb01;
+    char * dst_ptr = (char *)dst + i3*nb3 + i2*nb2 + i1*nb1;
+
+    for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
+        float value;
+
+        if (i0 < p0) {
+            // Left padding - reflect
+            value = *(const float *)(src0_ptr + (p0 - i0) * nb00);
+        } else if (i0 < ne0 - p1) {
+            // Middle - copy
+            value = *(const float *)(src0_ptr + (i0 - p0) * nb00);
+        } else {
+            // Right padding - reflect
+            int64_t src_idx = (ne0 - p1 - p0) - (p1 + 1 - (ne0 - i0)) - 1;
+            value = *(const float *)(src0_ptr + src_idx * nb00);
+        }
+
+        *(float *)(dst_ptr + i0 * nb0) = value;
+    }
+}
+
+void ggml_cuda_op_pad_reflect_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int32_t * opts = (const int32_t *) dst->op_params;
+    const int p0 = opts[0];
+    const int p1 = opts[1];
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t ne0 = dst->ne[0];
+
+    GGML_ASSERT(ne0 == ne00 + p0 + p1);
+
+    const dim3 block_dims(CUDA_PAD_REFLECT_1D_BLOCK_SIZE, 1, 1);
+    const dim3 grid_dims(ne01, ne02, ne03);
+
+    pad_reflect_1d_kernel_f32<<<grid_dims, block_dims, 0, stream>>>(
+        src0->data, dst->data,
+        ne0, ne00, ne01, ne02, ne03,
+        src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+        dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+        p0, p1
+    );
+}

ggml/src/ggml-cuda/pad_reflect_1d.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_PAD_REFLECT_1D_BLOCK_SIZE 256
+
+void ggml_cuda_op_pad_reflect_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/vecdotq.cuh

@@ -28,7 +28,58 @@ static __device__ __forceinline__ int get_int_b4(const void * x, const int & i32
     return ((const int *) x)[i32]; // assume at least 4 byte alignment
 }
 
+// q4 contains 8 indices with 4 bit each.
+// This function selects those bytes from table that are at those indices and returns them as int2.
+// The first int contains the bytes with even indices in q4, the second int contains the bytes with odd indices in q4.
 static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4, const int8_t * table) {
+#if defined(GGML_USE_HIP)
+    // Load the 16-byte table into four 32-bit unsigned integers.
+    const uint32_t *values = (const uint32_t *)table;
+
+    const uint32_t q_even = q4;
+    const uint32_t q_odd  = (q4 >> 4);
+
+    // Perform lookups in the lower half of the table (indices 0-7).
+    uint32_t v_even_low = __builtin_amdgcn_perm(values[1], values[0], q_even & 0x07070707);
+    uint32_t v_odd_low = __builtin_amdgcn_perm(values[1], values[0], q_odd & 0x07070707);
+
+    // Perform lookups in the upper half of the table (indices 8-15).
+    uint32_t v_even_high = __builtin_amdgcn_perm(values[3], values[2], q_even & 0x07070707);
+    uint32_t v_odd_high = __builtin_amdgcn_perm(values[3], values[2], q_odd & 0x07070707);
+
+    // Select between the low and high results based on the MSB of each index nibble.
+    uint32_t mask_even = 0x03020100 | ((q_even & 0x08080808) >> 1);
+    uint32_t res_x = __builtin_amdgcn_perm(v_even_high, v_even_low, mask_even);
+    uint32_t mask_odd = 0x03020100 | ((q_odd & 0x08080808) >> 1);
+    uint32_t res_y = __builtin_amdgcn_perm(v_odd_high, v_odd_low, mask_odd);
+
+    return make_int2(res_x, res_y);
+#elif !defined(GGML_USE_MUSA)
+    // CUDA does not have an instruction for selecting bytes with 4 bit indices.
+    // However, __byte_perm is an instruction that selects bytes with 3 bit indices that can be used instead.
+    const uint32_t * table32 = (const uint32_t *) table;
+
+    // __byte_perm selects bytes based on the lower 16 bits in its third argument.
+    // Therefore, do 2 iterations over the 32 bits in q4 with 0 and 16 shift.
+    // To handle the fourth bit, first call _byte_perm both for the low and the high 64 bit of table, using the low 3 bits.
+    // Then, call __byte_perm again to select from the low and high bytes based on the fourth bit.
+    uint32_t tmp[2];
+    const uint32_t low_high_selection_indices = (0x32103210 | ((q4 & 0x88888888) >> 1));
+#pragma unroll
+    for (uint32_t i = 0; i < 2; ++i) {
+        const uint32_t shift = 16 * i;
+
+        const uint32_t low  = __byte_perm(table32[0], table32[1], q4 >> shift);
+        const uint32_t high = __byte_perm(table32[2], table32[3], q4 >> shift);
+        tmp[i] = __byte_perm(low, high, low_high_selection_indices >> shift);
+    }
+
+    // tmp contains the bytes from tyble in the same order as the 4 bit indices in q4.
+    // However, for the result we need ints with all even/odd 4 bit indices in q4.
+    // Therefore, 2 more calls to __byte_perm to put the bytes in the correct order.
+    return make_int2(__byte_perm(tmp[0], tmp[1], 0x6420), __byte_perm(tmp[0], tmp[1], 0x7531));
+#else
+    // Generic implementation.
     const int      q0_32  = (q4 >> 0) & 0x0F0F0F0F;
     const int8_t * q0_8   = (const int8_t *) &q0_32;
     const char4    val0_8 = make_char4(
@@ -40,6 +91,7 @@ static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4, con
         table[q1_8[0]], table[q1_8[1]], table[q1_8[2]], table[q1_8[3]]);
 
     return make_int2(*((const int *) &val0_8), *((const int *) &val1_8));
+#endif
 }
 
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called

ggml/src/ggml-cuda/vendors/hip.h

@@ -22,7 +22,10 @@
 #define CU_MEM_ACCESS_FLAGS_PROT_READWRITE hipMemAccessFlagsProtReadWrite
 #define CU_CHECK(fn) {hipError_t err = fn; if(err != hipSuccess) { GGML_ABORT("HipVMM Failure: %s\n", hipGetErrorString(err)); }}
 #define __shfl_sync(mask, var, laneMask, width) __shfl(var, laneMask, width)
+#define __shfl_up_sync(mask, var, laneMask, width) __shfl_up(var, laneMask, width)
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define __all_sync(mask, var) __all(var)
+#define __any_sync(mask, var) __any(var)
 #define cublasCreate hipblasCreate
 #define cublasDestroy hipblasDestroy
 #define cublasGemmEx hipblasGemmEx

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

@@ -249,6 +249,7 @@ typedef struct {
     uint64_t nb33;
     int32_t  ne1;
     int32_t  ne2;
+    int32_t  ne3;
     float    scale;
     float    max_bias;
     float    m0;
@@ -257,6 +258,11 @@ typedef struct {
     float    logit_softcap;
 } ggml_metal_kargs_flash_attn_ext;
 
+typedef struct {
+    int32_t  nrows;
+    int32_t  ne20;
+} ggml_metal_kargs_flash_attn_ext_reduce;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne02;
@@ -320,40 +326,31 @@ typedef struct {
 } ggml_metal_kargs_mul_mv_ext;
 
 typedef struct {
+    int32_t  ne02;
     int32_t  ne10;
     int32_t  ne11;  // n_expert_used (bcast)
     uint64_t nb11;
     uint64_t nb12;
-    int32_t  neh11; // n_tokens
-    uint64_t nbh11;
+    int32_t  ne21; // n_tokens
     int32_t  ne20;  // n_expert_used
     uint64_t nb21;
 } ggml_metal_kargs_mul_mm_id_map0;
 
-typedef struct {
-    int32_t  ne20; // n_expert_used
-    int32_t  neh0;
-    int32_t  neh1;
-    uint64_t nbh1;
-    uint64_t nbh2;
-    int32_t  ne0;
-    uint64_t nb1;
-    uint64_t nb2;
-} ggml_metal_kargs_mul_mm_id_map1;
-
 typedef struct {
     int32_t  ne00;
     int32_t  ne02;
     uint64_t nb01;
     uint64_t nb02;
     uint64_t nb03;
-    int32_t  neh12;
-    uint64_t nbh10;
-    uint64_t nbh11;
-    uint64_t nbh12;
-    uint64_t nbh13;
-    int32_t  neh0;
-    int32_t  neh1;
+    int32_t  ne11;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne20;
+    int32_t  ne21;
+    int32_t  ne0;
+    int32_t  ne1;
     int16_t  r2;
     int16_t  r3;
 } ggml_metal_kargs_mul_mm_id;

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

@@ -93,35 +93,37 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
     if (ctx->mtl_device == nil) {
         ctx->mtl_device = MTLCreateSystemDefaultDevice();
 
-        ctx->has_simdgroup_reduction  = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
-        ctx->has_simdgroup_reduction |= [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
+        if (ctx->mtl_device) {
+            ctx->has_simdgroup_reduction  = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
+            ctx->has_simdgroup_reduction |= [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
 
-        ctx->has_simdgroup_mm = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
+            ctx->has_simdgroup_mm = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
 
 #if defined(GGML_METAL_HAS_RESIDENCY_SETS)
-        ctx->has_residency_sets = getenv("GGML_METAL_NO_RESIDENCY") == nil;
+            ctx->has_residency_sets = getenv("GGML_METAL_NO_RESIDENCY") == nil;
 #endif
 
-        ctx->has_bfloat  = [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
-        ctx->has_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
+            ctx->has_bfloat  = [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
+            ctx->has_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
 
 #if defined(GGML_METAL_USE_BF16)
-        ctx->use_bfloat = ctx->has_bfloat;
+            ctx->use_bfloat = ctx->has_bfloat;
 #else
-        ctx->use_bfloat = false;
+            ctx->use_bfloat = false;
 #endif
-        ctx->use_fusion = getenv("GGML_METAL_FUSION_DISABLE") == nil;
+            ctx->use_fusion = getenv("GGML_METAL_FUSION_DISABLE") == nil;
 
-        {
-            const char * val = getenv("GGML_METAL_FUSION_DEBUG");
-            ctx->debug_fusion = val ? atoi(val) : 0;
+            {
+                const char * val = getenv("GGML_METAL_FUSION_DEBUG");
+                ctx->debug_fusion = val ? atoi(val) : 0;
+            }
+
+            memset(ctx->fuse_cnt, 0, sizeof(ctx->fuse_cnt));
+
+            ctx->max_size = ctx->mtl_device.maxBufferLength;
+
+            strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1);
         }
-
-        memset(ctx->fuse_cnt, 0, sizeof(ctx->fuse_cnt));
-
-        ctx->max_size = ctx->mtl_device.maxBufferLength;
-
-        strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1);
     }
 
     ctx->mtl_device_ref_count++;
@@ -289,6 +291,10 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_MXFP4_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_5,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3,
     GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4,
@@ -396,8 +402,12 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP1_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_1,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_6,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_8,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_16,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F16,
@@ -443,6 +453,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,
     GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H40,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
@@ -452,6 +463,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H40,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96,
@@ -461,6 +473,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H40,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96,
@@ -470,6 +483,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H40,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96,
@@ -479,6 +493,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H40,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96,
@@ -488,6 +503,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H40,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96,
@@ -497,6 +513,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H40,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96,
@@ -506,6 +523,13 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H40,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,
@@ -555,6 +579,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_REDUCE,
     GGML_METAL_KERNEL_TYPE_SET_I32,
     GGML_METAL_KERNEL_TYPE_SET_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
@@ -1304,6 +1329,10 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,                 mul_mv_q5_1_f32,                 has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,                 mul_mv_q8_0_f32,                 has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_MXFP4_F32,                mul_mv_mxfp4_f32,                has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_2,         mul_mv_ext_f32_f32_r1_2,         has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_3,         mul_mv_ext_f32_f32_r1_3,         has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_4,         mul_mv_ext_f32_f32_r1_4,         has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_5,         mul_mv_ext_f32_f32_r1_5,         has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2,         mul_mv_ext_f16_f32_r1_2,         has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3,         mul_mv_ext_f16_f32_r1_3,         has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4,         mul_mv_ext_f16_f32_r1_4,         has_simdgroup_reduction);
@@ -1412,8 +1441,12 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,                mul_mm_iq1_m_f32,                has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,               mul_mm_iq4_nl_f32,               has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,               mul_mm_iq4_xs_f32,               has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16,              mul_mm_id_map0_f16,              has_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP1_F32,              mul_mm_id_map1_f32,              has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_1,       mul_mm_id_map0_f16_ne20_1,       has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_2,       mul_mm_id_map0_f16_ne20_2,       has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_4,       mul_mm_id_map0_f16_ne20_4,       has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_6,       mul_mm_id_map0_f16_ne20_6,       has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_8,       mul_mm_id_map0_f16_ne20_8,       has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_16,      mul_mm_id_map0_f16_ne20_16,      has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F16,               mul_mm_id_f32_f16,               has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F16,               mul_mm_id_f16_f16,               has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F16,              mul_mm_id_bf16_f16,              has_simdgroup_mm && use_bfloat);
@@ -1459,6 +1492,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,             argsort_f32_i32_asc,             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,            argsort_f32_i32_desc,            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,                  leaky_relu_f32,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H40,          flash_attn_ext_f16_h40,          has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,          flash_attn_ext_f16_h64,          has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,          flash_attn_ext_f16_h80,          has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,          flash_attn_ext_f16_h96,          has_simdgroup_mm);
@@ -1468,6 +1502,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128,  flash_attn_ext_f16_hk192_hv128,  has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,         flash_attn_ext_f16_h256,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512,  flash_attn_ext_f16_hk576_hv512,  has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H40,         flash_attn_ext_bf16_h40,         has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64,         flash_attn_ext_bf16_h64,         has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80,         flash_attn_ext_bf16_h80,         has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96,         flash_attn_ext_bf16_h96,         has_simdgroup_mm && use_bfloat);
@@ -1477,6 +1512,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128, flash_attn_ext_bf16_hk192_hv128, has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256,        flash_attn_ext_bf16_h256,        has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512, flash_attn_ext_bf16_hk576_hv512, has_simdgroup_mm && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H40,         flash_attn_ext_q4_0_h40,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64,         flash_attn_ext_q4_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80,         flash_attn_ext_q4_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96,         flash_attn_ext_q4_0_h96,         has_simdgroup_mm);
@@ -1486,6 +1522,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128, flash_attn_ext_q4_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256,        flash_attn_ext_q4_0_h256,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512, flash_attn_ext_q4_0_hk576_hv512, has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H40,         flash_attn_ext_q4_1_h40,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64,         flash_attn_ext_q4_1_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80,         flash_attn_ext_q4_1_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96,         flash_attn_ext_q4_1_h96,         has_simdgroup_mm);
@@ -1495,6 +1532,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128, flash_attn_ext_q4_1_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256,        flash_attn_ext_q4_1_h256,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512, flash_attn_ext_q4_1_hk576_hv512, has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H40,         flash_attn_ext_q5_0_h40,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64,         flash_attn_ext_q5_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80,         flash_attn_ext_q5_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96,         flash_attn_ext_q5_0_h96,         has_simdgroup_mm);
@@ -1504,6 +1542,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128, flash_attn_ext_q5_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256,        flash_attn_ext_q5_0_h256,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512, flash_attn_ext_q5_0_hk576_hv512, has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H40,         flash_attn_ext_q5_1_h40,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64,         flash_attn_ext_q5_1_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80,         flash_attn_ext_q5_1_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96,         flash_attn_ext_q5_1_h96,         has_simdgroup_mm);
@@ -1513,6 +1552,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128, flash_attn_ext_q5_1_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256,        flash_attn_ext_q5_1_h256,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512, flash_attn_ext_q5_1_hk576_hv512, has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H40,         flash_attn_ext_q8_0_h40,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64,         flash_attn_ext_q8_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80,         flash_attn_ext_q8_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96,         flash_attn_ext_q8_0_h96,         has_simdgroup_mm);
@@ -1522,6 +1562,13 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, flash_attn_ext_q8_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,        flash_attn_ext_q8_0_h256,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512, flash_attn_ext_q8_0_hk576_hv512, has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H40,      flash_attn_ext_vec_f16_h40,      has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40,     flash_attn_ext_vec_bf16_h40,     has_simdgroup_reduction && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40,     flash_attn_ext_vec_q4_0_h40,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40,     flash_attn_ext_vec_q4_1_h40,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40,     flash_attn_ext_vec_q5_0_h40,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40,     flash_attn_ext_vec_q5_1_h40,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40,     flash_attn_ext_vec_q8_0_h40,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,      flash_attn_ext_vec_f16_h64,      has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,     flash_attn_ext_vec_bf16_h64,     has_simdgroup_reduction && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,     flash_attn_ext_vec_q4_0_h64,     has_simdgroup_reduction);
@@ -1571,6 +1618,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512,    flash_attn_ext_vec_q5_0_hk576_hv512,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512,    flash_attn_ext_vec_q5_1_hk576_hv512,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512,    flash_attn_ext_vec_q8_0_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_REDUCE,           flash_attn_ext_reduce,           has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_F32,                         set_f32,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_I32,                         set_i32,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,                     cpy_f32_f32,                     true);
@@ -1846,7 +1894,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_ROPE:
             return true;
         case GGML_OP_IM2COL:
-            return op->src[1]->type == GGML_TYPE_F32 && (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32);
+            return ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32 && (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32);
         case GGML_OP_POOL_1D:
             return false;
         case GGML_OP_UPSCALE:
@@ -3347,15 +3395,16 @@ static int ggml_metal_encode_node(
 
                 // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                 // to the matrix-vector kernel
-                const int ne11_mm_min = 4;
+                const int ne11_mm_min = 8;
 
                 // first try to use small-batch mat-mv kernels
                 // these should be efficient for BS [2, ~8]
-                if (src1t == GGML_TYPE_F32 && (ne00%256 == 0) &&
+                if (src1t == GGML_TYPE_F32 && (ne00%128 == 0) &&
                     (
                      (
                       (
-                       src0t == GGML_TYPE_F16  || // TODO: helper function
+                       src0t == GGML_TYPE_F32  || // TODO: helper function
+                       src0t == GGML_TYPE_F16  ||
                        src0t == GGML_TYPE_Q4_0 ||
                        src0t == GGML_TYPE_Q4_1 ||
                        src0t == GGML_TYPE_Q5_0 ||
@@ -3383,7 +3432,17 @@ static int ggml_metal_encode_node(
                     //       values and there can be some tail effects when nsg is high. need to confirm this
                     //
                     const int nsg    = 2;                 // num simdgroups per threadgroup
-                    const int nxpsg  = ne11 < 3 ? 16 : 8; // num threads along row per simdgroup
+
+                    // num threads along row per simdgroup
+                    int nxpsg = 0;
+                    if (ne00 % 256 == 0 && ne11 < 3) {
+                        nxpsg = 16;
+                    } else if (ne00 % 128 == 0) {
+                        nxpsg = 8;
+                    } else {
+                        nxpsg = 4;
+                    }
+
                     const int nypsg  = 32/nxpsg;          // num threads along col per simdgroup (i.e. a simdgroup processes that many src0 rows at a time)
                     const int r0ptg  = nypsg*nsg;         // num src0 rows per threadgroup
                           int r1ptg  = 4;                 // num src1 rows per threadgroup
@@ -3406,6 +3465,14 @@ static int ggml_metal_encode_node(
                     id<MTLComputePipelineState> pipeline = nil;
 
                     switch (src0->type) {
+                        case GGML_TYPE_F32:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F32_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
                         case GGML_TYPE_F16:
                             switch (r1ptg) {
                                 case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2].pipeline; break;
@@ -3560,7 +3627,7 @@ static int ggml_metal_encode_node(
                         case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32   ].pipeline; break;
                         case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32   ].pipeline; break;
                         case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32   ].pipeline; break;
-                        case GGML_TYPE_MXFP4:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_MXFP4_F32   ].pipeline; break;
+                        case GGML_TYPE_MXFP4:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_MXFP4_F32  ].pipeline; break;
                         case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32   ].pipeline; break;
                         case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32   ].pipeline; break;
                         case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32   ].pipeline; break;
@@ -3878,38 +3945,6 @@ static int ggml_metal_encode_node(
                         default: break;
                     }
 
-                    const int64_t neh10 = ne10; // n_embd
-                    const int64_t neh11 = ne21; // n_tokens
-                    const int64_t neh12 = ne02; // n_expert
-
-                    const uint64_t nbh10 = ggml_type_size(GGML_TYPE_F16);
-                    const uint64_t nbh11 = nbh10*neh10;
-                    const uint64_t nbh12 = nbh11*neh11;
-                    const uint64_t nbh13 = nbh12*neh12;
-
-                    const size_t s_src1 = ggml_type_size(GGML_TYPE_F16)*neh10*neh11*neh12;
-                    id<MTLBuffer> h_src1 = ggml_metal_mem_pool_alloc(mem_pool, s_src1);
-                    if (!h_src1) {
-                        GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_src1);
-                        return 0;
-                    }
-
-                    const int64_t neh0 = ne0;
-                    const int64_t neh1 = ne21;
-                    const int64_t neh2 = ne02;
-
-                    const uint64_t nbh0 = ggml_type_size(GGML_TYPE_F32);
-                    const uint64_t nbh1 = nbh0*neh0;
-                    const uint64_t nbh2 = nbh1*neh1;
-                  //const uint64_t nbh3 = nbh2*neh2;
-
-                    const size_t s_dst = ggml_type_size(GGML_TYPE_F32)*neh0*neh1*neh2;
-                    id<MTLBuffer> h_dst = ggml_metal_mem_pool_alloc(mem_pool, s_dst);
-                    if (!h_dst) {
-                        GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_dst);
-                        return 0;
-                    }
-
                     // tokens per expert
                     const size_t s_tpe = ggml_type_size(GGML_TYPE_I32)*ne02;
                     id<MTLBuffer> h_tpe = ggml_metal_mem_pool_alloc(mem_pool, s_tpe);
@@ -3919,8 +3954,8 @@ static int ggml_metal_encode_node(
                     }
 
                     // id map
-                    // [n_expert_used, n_tokens]
-                    const size_t s_ids = ggml_type_size(GGML_TYPE_I32)*ne20*ne21;
+                    // [n_tokens, n_expert]
+                    const size_t s_ids = ggml_type_size(GGML_TYPE_I32)*ne21*ne02;
                     id<MTLBuffer> h_ids = ggml_metal_mem_pool_alloc(mem_pool, s_ids);
                     if (!h_ids) {
                         GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_ids);
@@ -3928,32 +3963,45 @@ static int ggml_metal_encode_node(
                     }
 
                     {
-                        const int nth = MIN(1024, ne10/4);
-
                         ggml_metal_kargs_mul_mm_id_map0 args = {
+                            ne02,
                             ne10,
-                            ne11,  // n_expert_used (bcast)
+                            ne11, // n_expert_used (bcast)
                             nb11,
                             nb12,
-                            neh11, // n_tokens
-                            nbh11,
-                            ne20,  // n_expert_used
+                            ne21, // n_tokens
+                            ne20, // n_expert_used
                             nb21,
                         };
 
                         id<MTLComputePipelineState> pipeline = nil;
 
-                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16].pipeline;
+                        pipeline = nil;
+
+                        switch (ne20) {
+                            case 1:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_1 ].pipeline; break;
+                            case 2:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_2 ].pipeline; break;
+                            case 4:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_4 ].pipeline; break;
+                            case 6:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_6 ].pipeline; break;
+                            case 8:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_8 ].pipeline; break;
+                            case 16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16_NE20_16].pipeline; break;
+                            default: GGML_ABORT("missing specialization for ne20 = %d", (int) ne20);
+                        }
+
+                        GGML_ASSERT(ne02 <= (int) pipeline.maxTotalThreadsPerThreadgroup);
+
+                        const size_t smem = ne02*ne20*sizeof(uint16_t);
+
+                        GGML_ASSERT(smem <= device.maxThreadgroupMemoryLength);
 
                         [encoder setComputePipelineState:pipeline];
                         [encoder setBytes:&args    length:sizeof(args) atIndex:0];
-                        [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
-                        [encoder setBuffer:id_src2 offset:offs_src2    atIndex:2];
-                        [encoder setBuffer: h_src1 offset:0            atIndex:3];
-                        [encoder setBuffer: h_tpe  offset:0            atIndex:4];
-                        [encoder setBuffer: h_ids  offset:0            atIndex:5];
+                        [encoder setBuffer:id_src2 offset:offs_src2    atIndex:1];
+                        [encoder setBuffer: h_tpe  offset:0            atIndex:2];
+                        [encoder setBuffer: h_ids  offset:0            atIndex:3];
+                        [encoder setThreadgroupMemoryLength:smem atIndex:0];
 
-                        [encoder dispatchThreadgroups:MTLSizeMake(ne02, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        [encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(ne02, 1, 1)];
                     }
 
                     {
@@ -3992,13 +4040,15 @@ static int ggml_metal_encode_node(
                             /*.nb01  =*/ nb01,
                             /*.nb02  =*/ nb02,
                             /*.nb03  =*/ nb03,
-                            /*.neh12 =*/ neh12,
-                            /*.nbh10 =*/ nbh10,
-                            /*.nbh11 =*/ nbh11,
-                            /*.nbh12 =*/ nbh12,
-                            /*.nbh13 =*/ nbh13,
-                            /*.neh0  =*/ neh0,
-                            /*.neh1  =*/ neh1,
+                            /*.ne11  =*/ ne11, // n_expert_used (bcast)
+                            /*.nb10  =*/ nb10,
+                            /*.nb11  =*/ nb11,
+                            /*.nb12  =*/ nb12,
+                            /*.nb13  =*/ nb13,
+                            /*.ne20  =*/ ne20, // n_expert_used
+                            /*.ne21  =*/ ne21, // n_tokens
+                            /*.ne0   =*/ ne0,
+                            /*.ne1   =*/ ne1,
                             /*.r2    =*/ r2,
                             /*.r3    =*/ r3,
                         };
@@ -4006,42 +4056,14 @@ static int ggml_metal_encode_node(
                         [encoder setComputePipelineState:pipeline];
                         [encoder setBytes:&args    length:sizeof(args) atIndex:0];
                         [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
-                        [encoder setBuffer: h_src1 offset:0            atIndex:2];
+                        [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
                         [encoder setBuffer: h_tpe  offset:0            atIndex:3];
-                        [encoder setBuffer: h_dst  offset:0            atIndex:4];
+                        [encoder setBuffer: h_ids  offset:0            atIndex:4];
+                        [encoder setBuffer:id_dst  offset:offs_dst     atIndex:5];
 
                         [encoder setThreadgroupMemoryLength:8192 atIndex:0];
                         [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, ne02) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                     }
-
-                    {
-                        GGML_ASSERT(ne0 % 4 == 0);
-
-                        const int nth = MIN(1024, ne0/4);
-
-                        ggml_metal_kargs_mul_mm_id_map1 args = {
-                            ne20, // n_expert_used
-                            neh0,
-                            neh1,
-                            nbh1,
-                            nbh2,
-                            ne0,
-                            nb1,
-                            nb2,
-                        };
-
-                        id<MTLComputePipelineState> pipeline = nil;
-
-                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP1_F32].pipeline;
-
-                        [encoder setComputePipelineState:pipeline];
-                        [encoder setBytes:&args   length:sizeof(args) atIndex:0];
-                        [encoder setBuffer: h_dst offset:0            atIndex:1];
-                        [encoder setBuffer: h_ids offset:0            atIndex:2];
-                        [encoder setBuffer:id_dst offset:offs_dst     atIndex:3];
-
-                        [encoder dispatchThreadgroups:MTLSizeMake(ne20, ne21, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                    }
                 } else {
                     id<MTLComputePipelineState> pipeline = nil;
 
@@ -4701,7 +4723,6 @@ static int ggml_metal_encode_node(
             } break;
         case GGML_OP_IM2COL:
             {
-                GGML_ASSERT(ggml_is_contiguous(src0));
                 GGML_ASSERT(ggml_is_contiguous(src1));
                 GGML_ASSERT(src1->type == GGML_TYPE_F32);
                 GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
@@ -5130,6 +5151,7 @@ static int ggml_metal_encode_node(
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
+                                        case 40:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H40 ].pipeline; break;
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
                                         case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break;
                                         case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
@@ -5154,6 +5176,7 @@ static int ggml_metal_encode_node(
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
+                                        case 40:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H40 ].pipeline; break;
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64 ].pipeline; break;
                                         case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80 ].pipeline; break;
                                         case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96 ].pipeline; break;
@@ -5178,6 +5201,7 @@ static int ggml_metal_encode_node(
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
+                                        case 40:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H40 ].pipeline; break;
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64 ].pipeline; break;
                                         case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80 ].pipeline; break;
                                         case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96 ].pipeline; break;
@@ -5202,6 +5226,7 @@ static int ggml_metal_encode_node(
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
+                                        case 40:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H40 ].pipeline; break;
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64 ].pipeline; break;
                                         case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80 ].pipeline; break;
                                         case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96 ].pipeline; break;
@@ -5226,6 +5251,7 @@ static int ggml_metal_encode_node(
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
+                                        case 40:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H40 ].pipeline; break;
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64 ].pipeline; break;
                                         case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80 ].pipeline; break;
                                         case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96 ].pipeline; break;
@@ -5250,6 +5276,7 @@ static int ggml_metal_encode_node(
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
+                                        case 40:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H40 ].pipeline; break;
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64 ].pipeline; break;
                                         case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80 ].pipeline; break;
                                         case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96 ].pipeline; break;
@@ -5274,6 +5301,7 @@ static int ggml_metal_encode_node(
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
+                                        case 40:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H40 ].pipeline; break;
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64 ].pipeline; break;
                                         case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80 ].pipeline; break;
                                         case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96 ].pipeline; break;
@@ -5301,6 +5329,24 @@ static int ggml_metal_encode_node(
                     use_vec_kernel = true;
 
                     switch (ne00) {
+                        case 40:
+                            {
+                                switch (src1->type) {
+                                    case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H40].pipeline; break;
+                                    case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40].pipeline; break;
+                                    case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40].pipeline; break;
+                                    case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40].pipeline; break;
+                                    case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40].pipeline; break;
+                                    case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40].pipeline; break;
+                                    case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40].pipeline; break;
+                                    default:
+                                        {
+                                            GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
+                                            GGML_LOG_ERROR("add template specialization for this type\n");
+                                            GGML_ABORT("add template specialization for this type");
+                                        }
+                                }
+                            } break;
                         case 64:
                             {
                                 switch (src1->type) {
@@ -5465,6 +5511,7 @@ static int ggml_metal_encode_node(
                     /*.nb33          =*/ nb33,
                     /*.ne1           =*/ ne1,
                     /*.ne2           =*/ ne2,
+                    /*.ne3           =*/ ne3,
                     /*.scale         =*/ scale,
                     /*.max_bias      =*/ max_bias,
                     /*.m0            =*/ m0,
@@ -5488,7 +5535,6 @@ static int ggml_metal_encode_node(
                 } else {
                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:5];
                 }
-                [encoder setBuffer:id_dst offset:offs_dst       atIndex:6];
 
                 if (!use_vec_kernel) {
                     // half8x8 kernel
@@ -5514,7 +5560,7 @@ static int ggml_metal_encode_node(
 
                     while (true) {
                         const size_t smem = FATTN_SMEM(nsgmax);
-                        if (smem > device.maxThreadgroupMemoryLength) {
+                        if (smem > device.maxThreadgroupMemoryLength/2) {
                             break;
                         }
                         nsgmax *= 2;
@@ -5526,15 +5572,18 @@ static int ggml_metal_encode_node(
 
                     const size_t smem = FATTN_SMEM(nsg);
 
+                    [encoder setBuffer:id_dst offset:offs_dst atIndex:6];
+
                     //printf("smem: %zu, max: %zu, nsg = %d\n", smem, device.maxThreadgroupMemoryLength, (int) nsg);
                     GGML_ASSERT(smem <= device.maxThreadgroupMemoryLength);
                     [encoder setThreadgroupMemoryLength:smem atIndex:0];
-#undef FATTN_SMEM
                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
+#undef FATTN_SMEM
                 } else {
                     // half4x4 kernel
                     const int64_t nqptg = 1;  // queries per threadgroup    !! sync with kernel template arguments !!
                     const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
+                    const int64_t nkpsg = 1*ncpsg; // TODO: make adjustable
 
                     GGML_ASSERT(nqptg <= 32);
                     GGML_ASSERT(nqptg  % 1  == 0);
@@ -5544,15 +5593,17 @@ static int ggml_metal_encode_node(
                     // for each query, we load it as f16 in shared memory (ne00)
                     // and store the soft_max values and the mask
                     //
-                    // ne00*(nsg)
+                    // ne20*(nsg)
                     // each simdgroup has a full f32 head vector in shared mem to accumulate results
                     //
 #define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*ne20*(nsg))*(sizeof(float)/2), 16))
+//#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
                     while (true) {
                         const size_t smem = FATTN_SMEM(nsgmax);
-                        if (smem > device.maxThreadgroupMemoryLength) {
+                        // avoid using more than half of the threadgroup memory - can cause slow downs especially for large head sizes
+                        if (smem > device.maxThreadgroupMemoryLength/2) {
                             break;
                         }
                         nsgmax *= 2;
@@ -5560,7 +5611,7 @@ static int ggml_metal_encode_node(
                     nsgmax /= 2;
 
                     // simdgroups per threadgroup (a.k.a. warps)
-                    const int64_t nsgt = MAX(2, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)));
+                    const int64_t nsgt = MAX(2, MIN(nsgmax, MIN((ne11 + nkpsg - 1)/(nkpsg), (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)));
 
                     int64_t nsg = 1;
                     while (nsg <= nsgt) {
@@ -5568,13 +5619,74 @@ static int ggml_metal_encode_node(
                     }
                     nsg /= 2;
 
-                    const size_t smem = FATTN_SMEM(nsg);
+                    // workgroups
+                    // each workgroup handles nsg*nkpsg cache values
+                    uint16_t nwg = 1;
+                    if (4*nsg*nkpsg >= ne11) {
+                        const size_t smem = FATTN_SMEM(nsg);
 
-                    //printf("smem: %zu, max: %zu, nsg = %d\n", smem, device.maxThreadgroupMemoryLength, (int) nsg);
-                    GGML_ASSERT(smem <= device.maxThreadgroupMemoryLength);
-                    [encoder setThreadgroupMemoryLength:smem atIndex:0];
+                        //printf("smem: %zu, max: %zu, nsg = %d, nsgmax = %d\n", smem, device.maxThreadgroupMemoryLength, (int) nsg, (int) nsgmax);
+                        GGML_ASSERT(smem <= device.maxThreadgroupMemoryLength);
+
+                        // using 1 workgroup -> write the result directly into dst
+                        [encoder setBuffer:id_dst offset:offs_dst      atIndex:6];
+                        [encoder setBytes:&nwg length:sizeof(uint16_t) atIndex:7];
+
+                        [encoder setThreadgroupMemoryLength:smem atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
+                    } else {
+                        nwg = 32;
+                        nsg = MIN(4, nsg);
+
+                        const size_t smem = FATTN_SMEM(nsg);
+
+                        //printf("smem: %zu, max: %zu, nsg = %d, nsgmax = %d\n", smem, device.maxThreadgroupMemoryLength, (int) nsg, (int) nsgmax);
+                        GGML_ASSERT(smem <= device.maxThreadgroupMemoryLength);
+
+                        // sanity checks
+                        GGML_ASSERT(ne01*ne02*ne03 == ne1*ne2*ne3);
+                        GGML_ASSERT(ne1*ne2*ne3 <= (1u << 31));
+
+                        const int32_t nrows = ne1*ne2*ne3;
+
+                        // temp buffer for writing the results from each workgroup
+                        // - ne20: the size of the head vector
+                        // -  + 2: the S and M values for each intermediate result
+                        const size_t s_tmp = ggml_type_size(GGML_TYPE_F32)*(nrows*nwg*(ne20 + 2));
+                        id<MTLBuffer> h_tmp = ggml_metal_mem_pool_alloc(mem_pool, s_tmp);
+                        if (!h_tmp) {
+                            GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_tmp);
+                            return 0;
+                        }
+
+                        //printf("ne01 = %d, ne02 = %d, ne03 = %d, ne20 = %d\n", ne01, ne02, ne03, ne20);
+                        //printf("needed memory: %.3f MiB\n", (float) (ne01*ne02*ne03*ne20*sizeof(float))/1024.0f/1024.0f);
+
+                        [encoder setBuffer:h_tmp  offset:0             atIndex:6];
+                        [encoder setBytes:&nwg length:sizeof(uint16_t) atIndex:7];
+
+                        [encoder setThreadgroupMemoryLength:smem atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
+
+                        // reduce the results from the workgroups
+                        {
+                            ggml_metal_kargs_flash_attn_ext_reduce args0 = {
+                                nrows,
+                                ne20,
+                            };
+
+                            id<MTLComputePipelineState> pipeline0 = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_REDUCE].pipeline;
+
+                            [encoder setComputePipelineState:pipeline0];
+                            [encoder setBytes:&args0   length:sizeof(args0) atIndex:0];
+                            [encoder setBuffer:h_tmp   offset:0             atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst      atIndex:2];
+
+                            //printf("ne1 = %d, ne2 = %d, ne3 = %d, ne20 = %d\n", ne1, ne2, ne3, ne20);
+                            [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(32*32, 1, 1)];
+                        }
+                    }
 #undef FATTN_SMEM
-                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
                 }
             } break;
         case GGML_OP_DUP:

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

@@ -68,6 +68,11 @@ void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg)
     reg = (type4x4)(*src);
 }
 
+template <typename type4>
+void dequantize_f32_t4(device const float4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*src);
+}
+
 template <typename type4x4>
 void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
     reg = (type4x4)(*src);
@@ -974,9 +979,16 @@ kernel void kernel_mul(
     device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11 + args.o1[0];
     device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
 
-    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
-        const int i10 = i0%args.ne10;
-        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) * *((device float *)(src1_ptr + i10*args.nb10));
+    if (args.ne10 == 1) {
+        const float x = *((device float *)(src1_ptr));
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) * x;
+        }
+    } else {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const int i10 = i0%args.ne10;
+            *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) * *((device float *)(src1_ptr + i10*args.nb10));
+        }
     }
 }
 
@@ -1000,9 +1012,16 @@ kernel void kernel_div(
     device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11 + args.o1[0];
     device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
 
-    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
-        const int i10 = i0%args.ne10;
-        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) / *((device float *)(src1_ptr + i10*args.nb10));
+    if (args.ne10 == 1) {
+        const float x = 1.0f / *((device float *)(src1_ptr));
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) * x;
+        }
+    } else {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const int i10 = i0%args.ne10;
+            *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) / *((device float *)(src1_ptr + i10*args.nb10));
+        }
     }
 }
 
@@ -3001,7 +3020,6 @@ void kernel_mul_mv_ext_q4_f32_impl(
 #pragma unroll(r1ptg)
             for (short ir1 = 0; ir1 < r1ptg; ++ir1) {
                 sumf[ir1] += dot(lx[ch], y4[ir1][ch*nxpsg]);
-
             }
         }
 
@@ -3186,6 +3204,11 @@ kernel void kernel_mul_mv_ext_q4x4_f32_disp(
 typedef decltype(kernel_mul_mv_ext_q4_f32_disp  <2, block_q8_0, 32,  dequantize_q8_0_t4>) mul_mv_ext_q4_f32_t;
 typedef decltype(kernel_mul_mv_ext_q4x4_f32_disp<2, block_q4_K, 256, dequantize_q4_K>)    mul_mv_ext_q4x4_f32_t;
 
+template [[host_name("kernel_mul_mv_ext_f32_f32_r1_2")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, float4,       4,  dequantize_f32_t4>;
+template [[host_name("kernel_mul_mv_ext_f32_f32_r1_3")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, float4,       4,  dequantize_f32_t4>;
+template [[host_name("kernel_mul_mv_ext_f32_f32_r1_4")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, float4,       4,  dequantize_f32_t4>;
+template [[host_name("kernel_mul_mv_ext_f32_f32_r1_5")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, float4,       4,  dequantize_f32_t4>;
+
 template [[host_name("kernel_mul_mv_ext_f16_f32_r1_2")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, half4,        4,  dequantize_f16_t4>;
 template [[host_name("kernel_mul_mv_ext_f16_f32_r1_3")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, half4,        4,  dequantize_f16_t4>;
 template [[host_name("kernel_mul_mv_ext_f16_f32_r1_4")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, half4,        4,  dequantize_f16_t4>;
@@ -4663,6 +4686,7 @@ kernel void kernel_flash_attn_ext(
 
 typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64, 64>) flash_attn_ext_t;
 
+template [[host_name("kernel_flash_attn_ext_f16_h40" )]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  40,  40>;
 template [[host_name("kernel_flash_attn_ext_f16_h64" )]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  64,  64>;
 template [[host_name("kernel_flash_attn_ext_f16_h80" )]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  80,  80>;
 template [[host_name("kernel_flash_attn_ext_f16_h96" )]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  96,  96>;
@@ -4674,6 +4698,7 @@ template [[host_name("kernel_flash_attn_ext_f16_h256")]]         kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_hk576_hv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  576, 512>;
 
 #if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_bf16_h40" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
@@ -4685,6 +4710,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
 #endif
 
+template [[host_name("kernel_flash_attn_ext_q4_0_h40" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q4_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q4_0_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 96,  96>;
@@ -4695,6 +4721,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_hk192_hv128")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256, 256>;
 template [[host_name("kernel_flash_attn_ext_q4_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 576, 512>;
 
+template [[host_name("kernel_flash_attn_ext_q4_1_h40" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q4_1_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q4_1_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q4_1_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 96,  96>;
@@ -4705,6 +4732,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_hk192_hv128")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_1_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 256, 256>;
 template [[host_name("kernel_flash_attn_ext_q4_1_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 576, 512>;
 
+template [[host_name("kernel_flash_attn_ext_q5_0_h40" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q5_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q5_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q5_0_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 96,  96>;
@@ -4715,6 +4743,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_hk192_hv128")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 256, 256>;
 template [[host_name("kernel_flash_attn_ext_q5_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 576, 512>;
 
+template [[host_name("kernel_flash_attn_ext_q5_1_h40" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q5_1_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q5_1_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q5_1_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 96,  96>;
@@ -4725,6 +4754,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_hk192_hv128")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_1_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 256, 256>;
 template [[host_name("kernel_flash_attn_ext_q5_1_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 576, 512>;
 
+template [[host_name("kernel_flash_attn_ext_q8_0_h40" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q8_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q8_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q8_0_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 96,  96>;
@@ -4765,14 +4795,16 @@ kernel void kernel_flash_attn_ext_vec(
         device const char * mask,
         device const char * sinks,
         device       char * dst,
+        constant uint16_t & nwg,
         threadgroup  half * shmem_f16 [[threadgroup(0)]],
         uint3   tgpig[[threadgroup_position_in_grid]],
         ushort3   ntg[[threads_per_threadgroup]],
         ushort  tiisg[[thread_index_in_simdgroup]],
         ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
     const short nsg = ntg.y; // number of simdgroups
+    const short iwg = tgpig[2]%nwg;
 
-    const int iq3 = tgpig[2];
+    const int iq3 = tgpig[2]/nwg;
     const int iq2 = tgpig[1];
     const int iq1 = tgpig[0];
 
@@ -4851,7 +4883,7 @@ kernel void kernel_flash_attn_ext_vec(
 
         // loop over the KV cache
         // each simdgroup handles blocks of Q rows and C columns
-        for (int ic0 = 0; ic0 < args.ne11; ic0 += C*nsg) {
+        for (int ic0 = (int) iwg*C*nsg; ic0 < args.ne11; ic0 += (int) nwg*C*nsg) {
             const int ic = ic0 + C*sgitg;
             if (ic >= args.ne11) {
                 break;
@@ -4981,7 +5013,7 @@ kernel void kernel_flash_attn_ext_vec(
             }
         }
 
-        if (sinks != q && sgitg == 0) {
+        if (sinks != q && sgitg == 0 && iwg == 0) {
             const float m = M;
             const float s = tiisg == 0 ? ((device const float *) sinks)[iq2] : -FLT_MAX/2;
 
@@ -5090,14 +5122,25 @@ kernel void kernel_flash_attn_ext_vec(
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    device float4 * dst4 = (device float4 *) dst;
-
     // final rescale with 1/S and store to global memory
     if (sgitg == 0) {
-        const float S = ss[0];
+        const int64_t nrows = args.ne3*args.ne2*args.ne1;
+        const int64_t rid   = iq3*args.ne2*args.ne1 + iq2 + iq1*args.ne1;
 
+        device float4 * dst4 = (device float4 *) dst;
+        device float  * dst1 = (device float  *) dst + nrows*DV*nwg; // the S and M are stored after the results
+
+        const float S = nwg == 1 ? 1.0f/ss[0] : 1.0f;
+
+        // interleave the workgroup data
         for (short i = tiisg; i < DV4; i += NW) {
-            dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)iq1*args.ne1)*DV4 + i] = (float4) sr4[i]/S;
+            dst4[rid*DV4*nwg + nwg*i + iwg] = (float4) sr4[i]*S;
+        }
+
+        // store S and M
+        if (nwg > 1 && tiisg == 0) {
+            dst1[rid*(2*nwg) + 2*iwg + 0] = ss[0];
+            dst1[rid*(2*nwg) + 2*iwg + 1] = ss[1];
         }
     }
 }
@@ -5115,6 +5158,16 @@ kernel void kernel_flash_attn_ext_vec(
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 
+template [[host_name("kernel_flash_attn_ext_vec_f16_h40")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  40, 40, 8>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 40, 40, 8>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0,        8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 40, 40, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1,        8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 40, 40, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0,        8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 40, 40, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 40, 40, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 40, 40, 8>;
+
 template [[host_name("kernel_flash_attn_ext_vec_f16_h64")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  64, 64, 8>;
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_vec_bf16_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 64, 64, 8>;
@@ -5187,6 +5240,41 @@ template [[host_name("kernel_flash_attn_ext_vec_q8_0_hk576_hv512")]] kernel flas
 
 #undef FA_TYPES
 
+kernel void kernel_flash_attn_ext_reduce(
+        constant ggml_metal_kargs_flash_attn_ext_reduce & args,
+        device  const char * htmp,
+        device        char * dst,
+        uint   tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    const uint64_t rid = tgpig;
+
+    const short nwg = 32;
+    const short iwg = tiisg;
+    const short DV  = args.ne20;
+    const short DV4 = DV/4;
+
+    device const float4 * htmp4 = (device const float4 *) htmp + rid*DV4*nwg;
+    device const float  * ss    = (device const float  *) htmp + (uint64_t)args.nrows*DV*nwg;
+    device       float4 * dst4  = (device       float4 *) dst  + rid*DV4;
+
+    float S = ss[rid*(2*nwg) + 2*iwg + 0];
+    float M = ss[rid*(2*nwg) + 2*iwg + 1];
+
+    const float m  = simd_max(M);
+    const float ms = exp(M - m);
+
+    S = 1.0f/simd_sum(S*ms);
+
+    for (int i = sgitg; i < DV4; i += nwg) {
+        const float4 v = simd_sum(htmp4[i*nwg + iwg]*ms);
+
+        if (iwg == 0) {
+            dst4[i] = v*S;
+        }
+    }
+}
+
 template<typename T>
 kernel void kernel_set(
     constant ggml_metal_kargs_set & args,
@@ -7474,97 +7562,81 @@ kernel void kernel_mul_mm(
     }
 }
 
-template<typename T4>
+template<short ne20> // n_expert_used
 kernel void kernel_mul_mm_id_map0(
         constant ggml_metal_kargs_mul_mm_id_map0 & args,
-        device  const char * src1,
         device  const char * src2,
-        device        char * hsrc1,
         device        char * htpe,
         device        char * hids,
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
-    const int ide = tgpig[0]; // expert id
+        threadgroup   char * shmem [[threadgroup(0)]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    const short ide = tpitg; // expert id
 
-    int n_all = 0;
+    uint32_t n_all = 0;
 
-    device int32_t * ids_i32 = (device int32_t *) (hids);
+    device int32_t * ids_i32 = (device int32_t *) hids + ide*args.ne21;
 
-    for (int i21 = 0; i21 < args.neh11; i21++) { // n_tokens
-        device const int32_t * src2_i32 = (device const int32_t *) (src2 + i21*args.nb21);
+    for (int i21 = 0; i21 < args.ne21; i21 += ntg) { // n_tokens
+        if (i21 + tpitg < args.ne21) {
+            device const int32_t * src2_i32 = (device const int32_t *) (src2 + (i21 + tpitg)*args.nb21);
 
-        for (int i20 = 0; i20 < args.ne20; i20++) { // n_expert_used
-            if (src2_i32[i20] != ide) {
-                continue;
+            threadgroup uint16_t * sids = (threadgroup uint16_t *) shmem + tpitg*ne20;
+
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sids[i20] = src2_i32[i20];
             }
-
-            device const float4 *  src1_f32x4 = (device const float4 *) ( src1 + i21*args.nb12 + (i20%args.ne11)*args.nb11);
-            device       T4     * hsrc1_f32x4 = (device       T4     *) (hsrc1 + (ide*args.neh11 + n_all)*args.nbh11);
-
-            for (int64_t i00 = tpitg.x; i00 < args.ne10/4; i00 += ntg.x) {
-                hsrc1_f32x4[i00] = (T4) (src1_f32x4[i00]);
-            }
-
-            if (tpitg.x == 0) {
-                ids_i32[i21*args.ne20 + i20] = ide*args.neh11 + n_all;
-            }
-
-            ++n_all;
         }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (short t = 0; t < ntg; t++) {
+            if (i21 + t >= args.ne21) {
+                break;
+            }
+
+            threadgroup const uint16_t * sids = (threadgroup const uint16_t *) shmem + t*ne20;
+
+            short sel = 0;
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sel += (sids[i20] == ide)*(i20 + 1);
+            }
+
+            ids_i32[n_all] = (i21 + t)*ne20 + sel - 1;
+
+            n_all += sel > 0;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    if (tpitg.x == 0) {
-        device int32_t * tpe_i32 = (device int32_t *) (htpe);
-        tpe_i32[ide] = n_all;
-    }
+    device uint32_t * tpe_u32 = (device uint32_t *) (htpe);
+    tpe_u32[ide] = n_all;
 }
 
-typedef decltype(kernel_mul_mm_id_map0<half4>) kernel_mul_mm_id_map0_t;
+typedef decltype(kernel_mul_mm_id_map0<1>) kernel_mul_mm_id_map0_t;
 
-template [[host_name("kernel_mul_mm_id_map0_f16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<half4>;
-
-template<typename T>
-kernel void kernel_mul_mm_id_map1(
-        constant ggml_metal_kargs_mul_mm_id_map1 & args,
-        device  const char * hdst,
-        device  const char * hids,
-        device        char * dst,
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
-    const int i20 = tgpig[0]; // used expert
-    const int i21 = tgpig[1]; // token
-
-    device const int32_t * ids_i32    = (device const int32_t *) (hids);
-    device       float4  * dst_f32x4  = (device       float4  *) (dst + i20*args.nb1 + i21*args.nb2);
-
-    const int id = ids_i32[i21*args.ne20 + i20];
-
-    const int ide = id / args.neh1;
-    const int idt = id % args.neh1;
-
-    device const float4 * hdst_f32x4 = (device const float4 *) (hdst + idt*args.nbh1 + ide*args.nbh2);
-
-    for (int64_t i0 = tpitg.x; i0 < args.neh0/4; i0 += ntg.x) {
-        dst_f32x4[i0] = hdst_f32x4[i0];
-    }
-}
-
-typedef decltype(kernel_mul_mm_id_map1<float>) kernel_mul_mm_id_map1_t;
-
-template [[host_name("kernel_mul_mm_id_map1_f32")]] kernel kernel_mul_mm_id_map1_t kernel_mul_mm_id_map1<float>;
+template [[host_name("kernel_mul_mm_id_map0_f16_ne20_1" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<1>;
+template [[host_name("kernel_mul_mm_id_map0_f16_ne20_2" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<2>;
+template [[host_name("kernel_mul_mm_id_map0_f16_ne20_4" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<4>;
+template [[host_name("kernel_mul_mm_id_map0_f16_ne20_6" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<6>;
+template [[host_name("kernel_mul_mm_id_map0_f16_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
+template [[host_name("kernel_mul_mm_id_map0_f16_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
 
 template<typename T, typename T4x4, typename simdgroup_T8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
 kernel void kernel_mul_mm_id(
         constant ggml_metal_kargs_mul_mm_id & args,
         device const char * src0,
         device const char * src1,
-        device const char * tpe,
+        device const char * htpe,
+        device const char * hids,
         device       char * dst,
         threadgroup  char * shmem [[threadgroup(0)]],
         uint3  tgpig[[threadgroup_position_in_grid]],
         ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
         ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
     threadgroup T    * sa = (threadgroup T    *)(shmem);
@@ -7572,19 +7644,20 @@ kernel void kernel_mul_mm_id(
 
     const int r0 = tgpig.y;
     const int r1 = tgpig.x;
-    const int im = tgpig.z;
+    const int im = tgpig.z; // expert
 
-    device const int32_t * tpe_i32 = (device const int32_t *) (tpe);
+    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
+    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
 
-    const int neh1 = tpe_i32[im];
+    const int32_t neh1 = tpe_u32[im];
 
     if (r1*BLOCK_SIZE_N >= neh1) {
         return;
     }
 
     // if this block is of 64x32 shape or smaller
-    const short n_rows = (args.neh0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.neh0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    const short n_cols = (     neh1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (     neh1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
+    const short n_cols = (    neh1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (    neh1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
 
     // a thread shouldn't load data outside of the matrix
     const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
@@ -7600,20 +7673,23 @@ kernel void kernel_mul_mm_id(
 
     short il = (tiitg % THREAD_PER_ROW);
 
-    const int i12 = im%args.neh12;
-    const int i13 = im/args.neh12;
+    const int id = ids_i32[im*args.ne21 + r1*BLOCK_SIZE_N + thread_col];
 
-    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
     const short    offset1 = il/nl;
 
     device const block_q * x = (device const block_q *)(src0
         + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
 
-    device const half   * y = (device const half   *)(src1
-        + args.nbh13*i13
-        + args.nbh12*i12
-        + args.nbh11*(r1*BLOCK_SIZE_N + thread_col)
-        + args.nbh10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+    device const float   * y = (device const float   *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
 
     for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
         // load data and store to threadgroup memory
@@ -7629,7 +7705,7 @@ kernel void kernel_mul_mm_id(
             +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
         }
 
-        *(threadgroup half2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = *((device half2x4 *) y);
+        *(threadgroup half2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (half2x4)(*((device float2x4 *) y));
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
         x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
@@ -7665,43 +7741,38 @@ kernel void kernel_mul_mm_id(
         }
     }
 
-    if ((r0 + 1) * BLOCK_SIZE_M <= args.neh0 && (r1 + 1) * BLOCK_SIZE_N <= neh1) {
-        device float * C = (device float *) dst +
-            (BLOCK_SIZE_M * r0 + 32*(sgitg &  1)) + \
-            (BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.neh0 + im*args.neh1*args.neh0;
+    threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.neh0 * (i/4), args.neh0);
-        }
-    } else {
-        // block is smaller than 64x32, we should avoid writing data outside of the matrix
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                     + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
-        for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
+    threadgroup float * temp_str = ((threadgroup float *) shmem) \
+                                 + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
+
+    #pragma unroll(8)
+    for (short i = 0; i < 8; i++) {
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (short j = sgitg; j < n_cols; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1*BLOCK_SIZE_N + j];
+
+        const short ide = id % args.ne20;
+        const short idt = id / args.ne20;
+
+        device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float4 * D4 = (device float4 *) D;
+
+        threadgroup float  * C  = (threadgroup float  *) shmem + (j*BLOCK_SIZE_M);
+        threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+        int i = tiisg;
+        for (; i < n_rows/4; i += 32) {
+            *(D4 + i) = *(C4 + i);
         }
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        if (sgitg == 0) {
-            for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*args.neh0 + im*args.neh1*args.neh0;
-                device float4 * D4 = (device float4 *) D;
-
-                threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
-                threadgroup float4 * C4 = (threadgroup float4 *) C;
-
-                int i = 0;
-                for (; i < n_rows/4; i++) {
-                    *(D4 + i) = *(C4 + i);
-                }
-
-                i *= 4;
-                for (; i < n_rows; i++) {
-                    *(D + i) = *(C + i);
-                }
-            }
+        i = (4*(n_rows/4)) + tiisg;
+        for (; i < n_rows; i += 32) {
+            *(D + i) = *(C + i);
         }
     }
 }

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

@@ -420,9 +420,9 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_clamp;
     cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu, kernel_swiglu_oai, kernel_geglu_erf, kernel_geglu_quick,
               kernel_geglu_f16, kernel_reglu_f16, kernel_swiglu_f16, kernel_geglu_erf_f16, kernel_geglu_quick_f16;
-    cl_kernel kernel_norm;
+    cl_kernel kernel_norm, kernel_norm_mul_add;
     cl_kernel kernel_rms_norm, kernel_rms_norm_mul;
-    cl_kernel kernel_group_norm;
+    cl_kernel kernel_group_norm, kernel_group_norm_mul_add;
     cl_kernel kernel_diag_mask_inf, kernel_diag_mask_inf_8;
     cl_kernel kernel_soft_max, kernel_soft_max_4;
     cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
@@ -1161,7 +1161,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_norm =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_norm = clCreateKernel(backend_ctx->program_norm, "kernel_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_norm         = clCreateKernel(backend_ctx->program_norm, "kernel_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_norm_mul_add = clCreateKernel(backend_ctx->program_norm, "kernel_norm_mul_add", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1487,7 +1488,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         backend_ctx->program_group_norm =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_group_norm = clCreateKernel(backend_ctx->program_group_norm, "kernel_group_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_group_norm         = clCreateKernel(backend_ctx->program_group_norm, "kernel_group_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_group_norm_mul_add = clCreateKernel(backend_ctx->program_group_norm, "kernel_group_norm_mul_add", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -2498,12 +2500,47 @@ static bool ggml_opencl_can_fuse(const struct ggml_cgraph * cgraph, int node_idx
         if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
             return false;
         }
+    } else if (ops.size() == 3 && ops.begin()[0] == GGML_OP_NORM && ops.begin()[1] == GGML_OP_MUL && ops.begin()[2] == GGML_OP_ADD) {
+        const ggml_tensor *norm = cgraph->nodes[node_idx];
+        const ggml_tensor *mul  = cgraph->nodes[node_idx+1];
+        const ggml_tensor *add  = cgraph->nodes[node_idx+2];
+        const ggml_tensor *w    = mul->src[0] == norm ? mul->src[1] : mul->src[0];
+        const ggml_tensor *b    = add->src[0] == mul  ? add->src[1] : add->src[0];
+
+        // norm fusion only supports F32
+        if (norm->src[0]->type != GGML_TYPE_F32 || w->type != GGML_TYPE_F32 || b->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        if (norm->src[0]->ne[0] % 4 != 0) {
+            return false;
+        }
+
+        if (!ggml_is_contiguous(norm->src[0]) || !ggml_is_contiguous(w) || !ggml_is_contiguous(b)) {
+            return false;
+        }
+    } else if (ops.size() == 3 && ops.begin()[0] == GGML_OP_GROUP_NORM && ops.begin()[1] == GGML_OP_MUL && ops.begin()[2] == GGML_OP_ADD) {
+        const ggml_tensor *gn = cgraph->nodes[node_idx];
+        const ggml_tensor *mul = cgraph->nodes[node_idx+1];
+        const ggml_tensor *add = cgraph->nodes[node_idx+2];
+        const ggml_tensor *w   = mul->src[0] == gn ? mul->src[1] : mul->src[0];
+        const ggml_tensor *b   = add->src[0] == mul ? add->src[1] : add->src[0];
+
+        if (gn->src[0]->type != GGML_TYPE_F32 || w->type != GGML_TYPE_F32 || b->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        if (!ggml_is_contiguous(gn->src[0]) || !ggml_is_contiguous(w) || !ggml_is_contiguous(b)) {
+            return false;
+        }
     }
 
     return true;
 }
 
 static void ggml_opencl_op_rms_norm_fused(ggml_backend_t backend, ggml_tensor * rms_norm_tensor, ggml_tensor * mul_tensor);
+static void ggml_opencl_op_norm_fused(ggml_backend_t backend, ggml_tensor * norm_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor);
+static void ggml_opencl_op_group_norm_fused(ggml_backend_t backend, ggml_tensor * gn_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor);
 
 static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
@@ -2520,6 +2557,16 @@ static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggm
             continue;
         }
 
+        if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_NORM, GGML_OP_MUL, GGML_OP_ADD })) {
+            ggml_opencl_op_norm_fused(backend, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
+            i += 2;
+            continue;
+        }
+        if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_GROUP_NORM, GGML_OP_MUL, GGML_OP_ADD })) {
+            ggml_opencl_op_group_norm_fused(backend, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
+            i += 2;
+            continue;
+        }
         if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
             ggml_opencl_op_rms_norm_fused(backend, node, cgraph->nodes[i+1]);
             i++;
@@ -2647,8 +2694,9 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SOFT_MAX:
         case GGML_OP_NORM:
-        case GGML_OP_RMS_NORM:
             return true;
+        case GGML_OP_RMS_NORM:
+            return op->ne[0] % 4 == 0 && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_REPEAT:
             return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32; // Assuming F32 for now, can be expanded
         case GGML_OP_PAD:
@@ -5038,6 +5086,140 @@ static void ggml_opencl_op_rms_norm_fused(ggml_backend_t backend, ggml_tensor *
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
+static void ggml_opencl_op_norm_fused(ggml_backend_t backend, ggml_tensor * norm_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor) {
+    GGML_ASSERT(norm_tensor && mul_tensor && add_tensor);
+
+    const ggml_tensor * src0 = norm_tensor->src[0];
+    const ggml_tensor * src1 = mul_tensor->src[0] == norm_tensor ? mul_tensor->src[1] : mul_tensor->src[0];
+    const ggml_tensor * src2 = add_tensor->src[0] == mul_tensor ? add_tensor->src[1] : add_tensor->src[0];
+    const ggml_tensor * dst = add_tensor;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    float eps;
+    memcpy(&eps, norm_tensor->op_params, sizeof(float));
+
+    const int ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
+    const cl_ulong nb01 = src0->nb[1], nb02 = src0->nb[2], nb03 = src0->nb[3];
+    const int ne10 = src1->ne[0], ne11 = src1->ne[1], ne12 = src1->ne[2], ne13 = src1->ne[3];
+    const cl_ulong nb11 = src1->nb[1], nb12 = src1->nb[2], nb13 = src1->nb[3];
+    const int ne20 = src2->ne[0], ne21 = src2->ne[1], ne22 = src2->ne[2], ne23 = src2->ne[3];
+    const cl_ulong nb21 = src2->nb[1], nb22 = src2->nb[2], nb23 = src2->nb[3];
+    const cl_ulong nbd1 = dst->nb[1], nbd2 = dst->nb[2], nbd3 = dst->nb[3];
+
+    size_t sgs;
+    if (backend_ctx->gpu_family == ADRENO) sgs = 64;
+    else if (backend_ctx->gpu_family == INTEL) sgs = 32;
+    else GGML_ASSERT(false && "Unsupported GPU");
+
+    cl_kernel kernel = backend_ctx->kernel_norm_mul_add;
+
+    int nth = sgs;
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    while (nth < ne00/4 && nth < max_workgroup_size) nth *= 2;
+    nth = MIN(nth, max_workgroup_size);
+    nth = MIN(nth, ne00/4);
+
+    size_t gws[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t lws[] = {(size_t)nth, 1, 1};
+    size_t num_subgroups = (nth + sgs - 1) / sgs;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra2->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int), &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int), &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int), &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int), &ne20));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int), &ne21));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int), &ne22));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int), &ne23));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb21));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb22));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb23));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nbd1));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(cl_ulong), &nbd2));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(cl_ulong), &nbd3));
+    CL_CHECK(clSetKernelArg(kernel, 32, sizeof(float), &eps));
+    CL_CHECK(clSetKernelArg(kernel, 33, sizeof(cl_float2) * num_subgroups, NULL));
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, gws, lws, dst);
+}
+
+static void ggml_opencl_op_group_norm_fused(ggml_backend_t backend, ggml_tensor * gn_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor) {
+    GGML_ASSERT(gn_tensor && mul_tensor && add_tensor);
+
+    const ggml_tensor * src0 = gn_tensor->src[0];
+    const ggml_tensor * src1 = mul_tensor->src[0] == gn_tensor ? mul_tensor->src[1] : mul_tensor->src[0];
+    const ggml_tensor * src2 = add_tensor->src[0] == mul_tensor ? add_tensor->src[1] : add_tensor->src[0];
+    const ggml_tensor * dst = add_tensor;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    int groups;
+    float eps;
+    memcpy(&groups, gn_tensor->op_params, sizeof(int));
+    memcpy(&eps, (char *)gn_tensor->op_params + sizeof(int), sizeof(float));
+
+    cl_kernel kernel = backend_ctx->kernel_group_norm_mul_add;
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    int ne = ggml_nelements(src0);
+    int group_size = ne / groups;
+
+    size_t lws[] = { (size_t)MIN(max_workgroup_size, group_size) };
+    size_t gws[] = { (size_t)groups * lws[0] };
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra2->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &group_size));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(float), &eps));
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 1, gws, lws, dst);
+}
+
 static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);

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

@@ -70,3 +70,52 @@ kernel void kernel_group_norm(
         dst[j] *= scale;
     }
 }
+
+//------------------------------------------------------------------------------
+// group_norm_mul_add
+//------------------------------------------------------------------------------
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_group_norm_mul_add(
+        global float * src0, ulong offset0,
+        global float * src1, ulong offset1,
+        global float * src2, ulong offset2,
+        global float * dst, ulong offsetd,
+        int ne,
+        int group_size,
+        float eps
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    src1 = (global float *)((global char *)src1 + offset1);
+    src2 = (global float *)((global char *)src2 + offset2);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    int start = get_group_id(0) * group_size;
+    int end = start + group_size;
+    if (end > ne) {
+        end = ne;
+    }
+
+    float sum = 0.0f;
+    float sum_sq = 0.0f;
+
+    for (int j = start + get_local_id(0); j < end; j += get_local_size(0)) {
+        float val = src0[j];
+        sum += val;
+        sum_sq += val*val;
+    }
+
+    sum = sub_group_reduce_add(sum);
+    sum_sq = sub_group_reduce_add(sum_sq);
+
+    const float mean = sum / group_size;
+    const float var = sum_sq / group_size - mean * mean;
+    const float scale = rsqrt(var + eps);
+
+    for (int j = start + get_local_id(0); j < end; j += get_local_size(0)) {
+        dst[j] = ((src0[j] - mean) * scale) * src1[j] + src2[j];
+    }
+}

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

@@ -79,3 +79,83 @@ kernel void kernel_norm(
         y[i00] = y[i00] * scale;
     }
 }
+
+//------------------------------------------------------------------------------
+// norm_mul_add
+//------------------------------------------------------------------------------
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_norm_mul_add(
+        global char * src0_ptr, ulong src0_offset,
+        global char * src1_ptr, ulong src1_offset,
+        global char * src2_ptr, ulong src2_offset,
+        global char * dst_ptr,  ulong dst_offset,
+        int ne00, int ne01, int ne02, int ne03,
+        ulong nb01, ulong nb02, ulong nb03,
+        int ne10, int ne11, int ne12, int ne13,
+        ulong nb11, ulong nb12, ulong nb13,
+        int ne20, int ne21, int ne22, int ne23,
+        ulong nb21, ulong nb22, ulong nb23,
+        ulong nbd1, ulong nbd2, ulong nbd3,
+        float eps,
+        local float2 * sums
+) {
+    const int i03 = get_group_id(2);
+    const int i02 = get_group_id(1);
+    const int i01 = get_group_id(0);
+
+    global float4 * x = (global float4 *)(src0_ptr + src0_offset + i01*nb01 + i02*nb02 + i03*nb03);
+    global float4 * w = (global float4 *)(src1_ptr + src1_offset + (i01%ne11)*nb11 + (i02%ne12)*nb12 + (i03%ne13)*nb13);
+    global float4 * b = (global float4 *)(src2_ptr + src2_offset + (i01%ne21)*nb21 + (i02%ne22)*nb22 + (i03%ne23)*nb23);
+    global float4 * y = (global float4 *)(dst_ptr  + dst_offset  + i01*nbd1 + i02*nbd2 + i03*nbd3);
+
+    float p_sum = 0.0f;
+    float p_sum_sq = 0.0f;
+
+    const int n_chunks = ne00 / 4;
+    for (int i00 = get_local_id(0); i00 < n_chunks; i00 += get_local_size(0)) {
+        float4 val = x[i00];
+        p_sum += val.x + val.y + val.z + val.w;
+        p_sum_sq += dot(val, val);
+    }
+
+    p_sum = sub_group_reduce_add(p_sum);
+    p_sum_sq = sub_group_reduce_add(p_sum_sq);
+
+    if (get_sub_group_local_id() == 0) {
+        sums[get_sub_group_id()] = (float2)(p_sum, p_sum_sq);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (get_local_id(0) == 0) {
+        float sum = 0.0f;
+        float sum_sq = 0.0f;
+        for (uint i = 0; i < get_num_sub_groups(); ++i) {
+            float2 s = sums[i];
+            sum += s.x;
+            sum_sq += s.y;
+        }
+
+        const float inv_ne00 = 1.0f / (float)ne00;
+        const float mean = sum * inv_ne00;
+        const float variance = mad(-mean, mean, sum_sq * inv_ne00);
+
+        sums[0] = (float2)(mean, rsqrt(variance + eps));
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const float2 mean_scale = sums[0];
+    const float mean = mean_scale.x;
+    const float scale = mean_scale.y;
+    const float neg_mean_scale = -mean * scale;
+
+    for (int i00 = get_local_id(0); i00 < n_chunks; i00 += get_local_size(0)) {
+        const int w_idx = ne10 > 1 ? i00 : 0;
+        const int b_idx = ne20 > 1 ? i00 : 0;
+        const float4 norm_x = mad(x[i00], (float4)scale, (float4)neg_mean_scale);
+        y[i00] = mad(norm_x, w[w_idx], b[b_idx]);
+    }
+}

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

@@ -4364,11 +4364,12 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             return (op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32) && (op->type == op->src[0]->type);
 #endif
         case GGML_OP_NORM:
-        case GGML_OP_RMS_NORM:
             return true;
         case GGML_OP_L2_NORM:
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_RMS_NORM:
+            return ((op->src[0]->ne[0] % WARP_SIZE) == 0);
         case GGML_OP_SCALE:
             return true;
         case GGML_OP_CONT:
@@ -4391,10 +4392,11 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             return true;
         case GGML_OP_UPSCALE:
             return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
-        case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
+            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_POOL_2D:
         case GGML_OP_ACC:
         case GGML_OP_PAD:
         case GGML_OP_LEAKY_RELU:

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

@@ -1,20 +1,34 @@
 #version 450
 
 #extension GL_EXT_shader_16bit_storage : require
+#if ADD_RMS
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#endif
 
 #include "types.comp"
 #include "generic_binary_head.comp"
 
 const uint num_threads = 256;
 
+layout (binding = 3, std430) buffer PartialBuf {float partial_sums[];};
+
 layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
 
+#if ADD_RMS
+// XXX TODO this could be sized based on number of subgroups, but that't not considered a constant
+shared FLOAT_TYPE sumsh[num_threads];
+#endif
+
 void main() {
     uint idx = get_idx();
+    uint orig_idx = idx;
 
     // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
     const uint num_iter = 2;
 
+    FLOAT_TYPE sum_sq = 0;
+
     [[unroll]] for (uint i = 0; i < num_iter; ++i) {
         if (idx >= p.ne) {
             continue;
@@ -22,8 +36,34 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
+        FLOAT_TYPE sum = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]);
+        sum_sq += sum*sum;
+
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
 
         idx += num_threads;
     }
+
+#if ADD_RMS
+    if (p.param3 != 0) {
+        // reduce the sum within each subgroup, then across subgroups
+        const uint NumSubgroups = num_threads / gl_SubgroupSize;
+        sum_sq = subgroupAdd(sum_sq);
+        if (gl_SubgroupInvocationID == 0) {
+            sumsh[gl_SubgroupID] = sum_sq;
+        }
+        barrier();
+        [[unroll]] for (uint s = NumSubgroups / 2; s > 0; s >>= 1) {
+            if (gl_SubgroupID < s && gl_SubgroupInvocationID == 0) {
+                sum_sq += sumsh[gl_SubgroupID + s];
+                sumsh[gl_SubgroupID] = sum_sq;
+            }
+            barrier();
+        }
+
+        if (gl_SubgroupID == 0 && gl_SubgroupInvocationID == 0) {
+            partial_sums[orig_idx / (num_iter * num_threads)] = sum_sq;
+        }
+    }
+#endif
 }

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

@@ -0,0 +1,20 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+    data_d[i] = D_TYPE(exp(float(data_a[i])));
+}

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

@@ -9,6 +9,10 @@ layout (constant_id = 4) const uint32_t HSV = 32;
 layout (constant_id = 5) const uint32_t Clamp = 0;
 layout (constant_id = 6) const uint32_t D_split = 16;
 
+// Round up head sizes to a multiple of 16, for coopmat1/coopmat2 paths
+const uint32_t HSK_pad = (HSK + 15) & ~15;
+const uint32_t HSV_pad = (HSV + 15) & ~15;
+
 layout (push_constant) uniform parameter {
     uint32_t N;
     uint32_t KV;

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -46,14 +46,14 @@ const uint32_t MatBc = 16;
 shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
 shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x];
 
-const uint32_t qstride = HSK / 4 + 2; // in units of f16vec4
+const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
 shared f16vec4 Qf[Br * qstride];
 
 // Avoid padding for hsk==256 to make it fit in 48KB shmem.
 const uint32_t sfshstride = (HSK <= 128) ? (Br + 8) : Br;
 shared ACC_TYPE sfsh[Bc * sfshstride];
 
-const uint32_t kshstride = HSK / 4 + 2; // in units of f16vec4
+const uint32_t kshstride = HSK_pad / 4 + 2; // in units of f16vec4
 shared f16vec4 ksh[Bc * kshstride];
 
 shared float slope[Br];
@@ -74,6 +74,21 @@ void main() {
 
 #define tile_row(r) (row_tid * rows_per_thread + (r))
 
+    // Zero-initialize shared memory for Q/K when HSK is not a multiple of 16 (HSK_pad > HSK).
+    if ((HSK % 16) != 0) {
+        [[unroll]] for (uint i = 0; i < Br * qstride; i += gl_WorkGroupSize.x) {
+            if (i + tid < Br * qstride) {
+                Qf[i + tid] = f16vec4(0);
+            }
+        }
+        [[unroll]] for (uint i = 0; i < Bc * kshstride; i += gl_WorkGroupSize.x) {
+            if (i + tid < Bc * kshstride) {
+                ksh[i + tid] = f16vec4(0);
+            }
+        }
+        barrier();
+    }
+
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
     [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
@@ -151,14 +166,14 @@ void main() {
         }
         barrier();
 
-        // K * Q^T -> S^T: Bc x HSK * HSK x Br -> Bc x Br
+        // K * Q^T -> S^T: Bc x HSK_pad * HSK_pad x Br -> Bc x Br
         // Bc split across workgroup (four subgroups), loop over HSK in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
         // This is written transposed in order to allow for N being 8 if implementations need it
         coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
         coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat;
         coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
 
-        for (uint32_t d = 0; d < HSK / 16; ++d) {
+        for (uint32_t d = 0; d < HSK_pad / 16; ++d) {
             coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
 
             uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -104,16 +104,16 @@ void main() {
     tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
     tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
 
-    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseAccumulator> Q;
-    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA> Qf16;
+    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseAccumulator> Q;
+    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA> Qf16;
 
     uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
-    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK));
+    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK_pad));
 
-    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA>(Q);
+    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA>(Q);
     Qf16 *= float16_t(p.scale);
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
 
@@ -140,10 +140,10 @@ void main() {
 
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, HSK, Bc, gl_MatrixUseB> K_T;
+        coopmat<float16_t, gl_ScopeWorkgroup, HSK_pad, Bc, gl_MatrixUseB> K_T;
 
         uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
-        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK), tensorViewTranspose DECODEFUNC);
+        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK_pad), tensorViewTranspose DECODEFUNC);
         S = coopMatMulAdd(Qf16, K_T, S);
 
         if (p.logit_softcap != 0.0f) {
@@ -208,31 +208,31 @@ void main() {
         rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0);
         rowsum = coopMatMulAdd(P_A, One, rowsum);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV, gl_MatrixUseB> V;
+        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV_pad, gl_MatrixUseB> V;
         uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23;
-        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV) DECODEFUNC);
+        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV_pad) DECODEFUNC);
 
         L = eM*L + rowsum;
 
         // This is the "diagonal" matrix in the paper, but since we do componentwise
         // multiply rather than matrix multiply it has the diagonal element smeared
         // across the row
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> eMdiag;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> eMdiag;
 
         // resize eM by using smear/reduce
         coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
         // multiply with fp16 accumulation, then add to O.
-        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
         PV = coopMatMulAdd(P_A, V, PV);
 
-        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(PV);
+        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(PV);
     }
 
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
+        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
 
         uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
@@ -243,16 +243,16 @@ void main() {
         return;
     }
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Ldiag;
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> Ldiag;
 
     // resize L by using smear/reduce
     coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
     if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> S;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> S;
         coopMatPerElementNV(S, S, perElemOpGetSink, iq2);
 
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Mr;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> Mr;
 
         // resize M by using smear/reduce
         coopMatReduceNV(Mr, M, gl_CooperativeMatrixReduceRowNV, smearReduce);
@@ -285,7 +285,7 @@ void main() {
 
     uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
-    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
+    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
     if (p.gqa_ratio > 1) {
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
     } else {
@@ -295,6 +295,6 @@ void main() {
         // permute dimensions
         tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
 
-        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV), tensorViewPermute);
+        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV_pad), tensorViewPermute);
     }
 }

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

@@ -17,6 +17,9 @@
 #ifdef COOPMAT
 #extension GL_KHR_cooperative_matrix : enable
 #extension GL_KHR_memory_scope_semantics : enable
+#endif
+
+#if defined(COOPMAT) || defined(MUL_MAT_ID_USE_SUBGROUPS)
 #extension GL_KHR_shader_subgroup_basic : enable
 #extension GL_KHR_shader_subgroup_ballot : enable
 #endif
@@ -103,16 +106,79 @@ layout (constant_id = 10) const uint WARP = 32;
 shared FLOAT_TYPE buf_a[BM * SHMEM_STRIDE];
 shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE];
 
-#ifdef MUL_MAT_ID
-shared u16vec2 row_ids[4096];
-uint _ne1;
-#ifdef COOPMAT
-shared uint _ne1_sh;
-#endif
-#endif // MUL_MAT_ID
-
 #define NUM_WARPS (BLOCK_SIZE / WARP)
 
+#ifdef MUL_MAT_ID
+shared u16vec2 row_ids[BN];
+uint _ne1;
+
+#ifdef MUL_MAT_ID_USE_SUBGROUPS
+shared uvec4 ballots_sh[NUM_WARPS];
+
+void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
+    _ne1 = 0;
+    uint num_elements = p.nei1 * p.nei0;
+    uint nei0shift = findLSB(p.nei0);
+
+    uint ids[16];
+    uint iter = 0;
+
+    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
+        // prefetch up to 16 elements
+        if (iter == 0) {
+            [[unroll]] for (uint k = 0; k < 16; ++k) {
+                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
+                bool in_range = i < num_elements;
+                uint ii1;
+                if (nei0_is_pow2) {
+                    ii1 = i >> nei0shift;
+                } else {
+                    ii1 = i / p.nei0;
+                }
+                uint ii0 = i - ii1 * p.nei0;
+                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+            }
+        }
+        uint i = j + gl_LocalInvocationIndex;
+        bool in_range = i < num_elements;
+        uint ii1;
+        if (nei0_is_pow2) {
+            ii1 = i >> nei0shift;
+        } else {
+            ii1 = i / p.nei0;
+        }
+        uint ii0 = i - ii1 * p.nei0;
+        uint id = ids[iter++];
+        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
+
+        ballots_sh[gl_SubgroupID] = ballot;
+        barrier();
+
+        uint subgroup_base = 0;
+        uint total = 0;
+        for (uint k = 0; k < gl_NumSubgroups; ++k) {
+            if (k == gl_SubgroupID) {
+                subgroup_base = total;
+            }
+            total += subgroupBallotBitCount(ballots_sh[k]);
+        }
+        barrier();
+
+        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
+        if (in_range && id == expert_idx && _ne1 + idx >= ic * BN && _ne1 + idx < (ic + 1) * BN) {
+            row_ids[_ne1 + idx - ic * BN] = u16vec2(ii0, ii1);
+        }
+        _ne1 += total;
+        iter &= 15;
+        if (_ne1 >= (ic + 1) * BN) {
+            break;
+        }
+    }
+    barrier();
+}
+#endif // MUL_MAT_ID_USE_SUBGROUPS
+#endif // MUL_MAT_ID
+
 #ifdef COOPMAT
 shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
 #endif
@@ -177,51 +243,20 @@ void main() {
     const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;
 
 #ifdef MUL_MAT_ID
-#ifdef COOPMAT
-    // Spread the search across all elements in the first subgroup
-    if (gl_SubgroupID == 0) {
-        _ne1 = 0;
-        uint num_elements = p.nei1 * p.nei0;
-
-        uint ids[16];
-        uint iter = 0;
-
-        for (uint j = 0; j < num_elements; j += gl_SubgroupSize) {
-            // prefetch up to 16 elements
-            if (iter == 0) {
-                [[unroll]] for (uint k = 0; k < 16; ++k) {
-                    uint i = j + gl_SubgroupInvocationID + k*gl_SubgroupSize;
-                    bool in_range = i < num_elements;
-                    uint ii1 = i / p.nei0;
-                    uint ii0 = i % p.nei0;
-                    ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
-                }
-            }
-            uint i = j + gl_SubgroupInvocationID;
-            bool in_range = i < num_elements;
-            uint ii1 = i / p.nei0;
-            uint ii0 = i % p.nei0;
-            uint id = ids[iter++];
-            uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
-            uint idx = subgroupBallotExclusiveBitCount(ballot);
-            if (in_range && id == expert_idx) {
-                row_ids[_ne1 + idx] = u16vec2(ii0, ii1);
-            }
-            _ne1 += subgroupBallotBitCount(ballot);
-            iter &= 15;
-        }
-        _ne1_sh = _ne1;
+#ifdef MUL_MAT_ID_USE_SUBGROUPS
+    if (bitCount(p.nei0) == 1) {
+        load_row_ids(expert_idx, true, ic);
+    } else {
+        load_row_ids(expert_idx, false, ic);
     }
-
-    barrier();
-
-    _ne1 = _ne1_sh;
 #else
     _ne1 = 0;
-    for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
-        for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
+    for (uint ii1 = 0; ii1 < p.nei1 && _ne1 < (ic + 1) * BN; ii1++) {
+        for (uint ii0 = 0; ii0 < p.nei0 && _ne1 < (ic + 1) * BN; ii0++) {
             if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
-                row_ids[_ne1] = u16vec2(ii0, ii1);
+                if (_ne1 >= ic * BN) {
+                    row_ids[_ne1 - ic * BN] = u16vec2(ii0, ii1);
+                }
                 _ne1++;
             }
         }
@@ -767,7 +802,7 @@ void main() {
         [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
 #if LOAD_VEC_B == 8
 #ifdef MUL_MAT_ID
-            const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
+            const u16vec2 row_idx = row_ids[loadc_b + l];
             const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
 #else
             const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
@@ -783,7 +818,7 @@ void main() {
             buf_b[buf_idx + 7] = FLOAT_TYPE(data_b[idx][1].w);
 #elif LOAD_VEC_B == 4
 #ifdef MUL_MAT_ID
-            const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
+            const u16vec2 row_idx = row_ids[loadc_b + l];
             const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
 #else
             const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
@@ -802,7 +837,7 @@ void main() {
 #else
             const uint row_i = ic * BN + loadc_b + l;
             if (row_i < _ne1 && block + loadr_b < end_k) {
-                const u16vec2 row_idx = row_ids[row_i];
+                const u16vec2 row_idx = row_ids[loadc_b + l];
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
             } else {
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f);
@@ -873,7 +908,7 @@ void main() {
                 const uint row_i = dc + cm_col * TN + col + store_c;
                 if (row_i >= _ne1) break;
 
-                const u16vec2 row_idx = row_ids[row_i];
+                const u16vec2 row_idx = row_ids[row_i - ic * BN];
 
                 if (dr + cm_row * TM + store_r < p.M) {
                     data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
@@ -923,7 +958,7 @@ void main() {
                 const uint row_i = dc_warp + cc;
                 if (row_i >= _ne1) break;
 
-                const u16vec2 row_idx = row_ids[row_i];
+                const u16vec2 row_idx = row_ids[row_i - ic * BN];
 #endif // MUL_MAT_ID
                 [[unroll]] for (uint cr = 0; cr < TM; cr++) {
 #ifdef MUL_MAT_ID

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

@@ -19,6 +19,7 @@
 #endif
 
 #include "types.comp"
+#include "utils.comp"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
@@ -92,14 +93,15 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
 
-shared u16vec4 row_ids[4096];
+shared u16vec4 row_ids[BN];
 
 layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufB {
    B_TYPE b[];
 };
 
 uint _ne1;
-shared uint _ne1_sh;
+layout (constant_id = 5) const uint subgroup_size = 32;
+shared uvec4 ballots_sh[BLOCK_SIZE / subgroup_size];
 
 B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
@@ -109,7 +111,7 @@ B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const i
         return B_TYPE(0.0);
     }
 
-    const u16vec4 row_idx = row_ids[row_i];
+    const u16vec4 row_idx = row_ids[row_i & (BN - 1)];
     B_TYPE ret = data_b[row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + blockCoords[1]];
 
     return ret;
@@ -121,13 +123,74 @@ D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem
     uint dc = ic * BN + c;
 
     if (dr < p.M && dc < _ne1) {
-        uint row_i = dc;
+        uint row_i = c;
         const u16vec4 row_idx = row_ids[row_i];
         data_d[row_idx.y * p.batch_stride_d + row_idx.z * p.stride_d + dr] = elem;
     }
     return elem;
 }
 
+void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
+    _ne1 = 0;
+    uint num_elements = p.nei1 * p.nei0;
+    uint nei0shift = findLSB(p.nei0);
+
+    uint ids[16];
+    uint iter = 0;
+
+    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
+        // prefetch up to 16 elements
+        if (iter == 0) {
+            [[unroll]] for (uint k = 0; k < 16; ++k) {
+                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
+                bool in_range = i < num_elements;
+                uint ii1;
+                if (nei0_is_pow2) {
+                    ii1 = i >> nei0shift;
+                } else {
+                    ii1 = i / p.nei0;
+                }
+                uint ii0 = i - ii1 * p.nei0;
+                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+            }
+        }
+        uint i = j + gl_LocalInvocationIndex;
+        bool in_range = i < num_elements;
+        uint ii1;
+        if (nei0_is_pow2) {
+            ii1 = i >> nei0shift;
+        } else {
+            ii1 = i / p.nei0;
+        }
+        uint ii0 = i - ii1 * p.nei0;
+        uint id = ids[iter++];
+        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
+
+        ballots_sh[gl_SubgroupID] = ballot;
+        barrier();
+
+        uint subgroup_base = 0;
+        uint total = 0;
+        for (uint k = 0; k < gl_NumSubgroups; ++k) {
+            if (k == gl_SubgroupID) {
+                subgroup_base = total;
+            }
+            total += subgroupBallotBitCount(ballots_sh[k]);
+        }
+        barrier();
+
+        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
+        if (in_range && id == expert_idx && _ne1 + idx >= ic * BN && _ne1 + idx < (ic + 1) * BN) {
+            row_ids[_ne1 + idx - ic * BN] = u16vec4(fastmod(ii0, p.ne11), ii1, ii0, 0);
+        }
+        _ne1 += total;
+        iter &= 15;
+        if (_ne1 >= (ic + 1) * BN) {
+            break;
+        }
+    }
+    barrier();
+}
 #endif
 
 void main() {
@@ -157,45 +220,12 @@ void main() {
     const uint ic = gl_WorkGroupID.y;
 
 #ifdef MUL_MAT_ID
-    // Spread the search across all elements in the first subgroup
-    if (gl_SubgroupID == 0) {
-        _ne1 = 0;
-        uint num_elements = p.nei1 * p.nei0;
-
-        uint ids[16];
-        uint iter = 0;
-
-        for (uint j = 0; j < num_elements; j += gl_SubgroupSize) {
-            // prefetch up to 16 elements
-            if (iter == 0) {
-                [[unroll]] for (uint k = 0; k < 16; ++k) {
-                    uint i = j + gl_SubgroupInvocationID + k*gl_SubgroupSize;
-                    bool in_range = i < num_elements;
-                    uint ii1 = i / p.nei0;
-                    uint ii0 = i % p.nei0;
-                    ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
-                }
-            }
-            uint i = j + gl_SubgroupInvocationID;
-            bool in_range = i < num_elements;
-            uint ii1 = i / p.nei0;
-            uint ii0 = i % p.nei0;
-            uint id = ids[iter++];
-            uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
-            uint idx = subgroupBallotExclusiveBitCount(ballot);
-            if (in_range && id == expert_idx) {
-                row_ids[_ne1 + idx] = u16vec4(ii0 % p.ne11, ii1, ii0, 0);
-            }
-            _ne1 += subgroupBallotBitCount(ballot);
-            iter &= 15;
-        }
-        _ne1_sh = _ne1;
+    if (bitCount(p.nei0) == 1) {
+        load_row_ids(expert_idx, true, ic);
+    } else {
+        load_row_ids(expert_idx, false, ic);
     }
 
-    barrier();
-
-    _ne1 = _ne1_sh;
-
     // Workgroup has no work
     if (ic * BN >= _ne1) return;
 #endif

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

@@ -3,6 +3,10 @@
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_nonuniform_qualifier : enable
 #extension GL_EXT_control_flow_attributes : require
+#if ADD_RMS
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#endif
 
 #include "rte.comp"
 #include "types.comp"
@@ -14,11 +18,18 @@ layout (push_constant) uniform parameter2
     uint ne20; uint ne21; uint ne22; uint ne23;
 
     // strides for srcs+dst
-    uint nb[8][4];
+    uint nb[12][4];
+
+    uint rms_partials;
 } p;
 
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
-layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
+// Workaround for MoltenVK Bug, see https://github.com/ggml-org/llama.cpp/issues/15498
+// layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
+// layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
+layout (binding = 0) buffer A {A_TYPE data_a[];} a[];
+layout (binding = 0) buffer D {D_TYPE data_d[];} d[];
+
+layout (binding = 0, std430) buffer PartialBuf {float partial_sums[];} partials[];
 
 layout(constant_id = 0) const uint num_srcs = 2;
 
@@ -42,14 +53,22 @@ const uint num_threads = 256;
 
 layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
 
+#if ADD_RMS
+// XXX TODO this could be sized based on number of subgroups, but that't not considered a constant
+shared FLOAT_TYPE sumsh[num_threads];
+#endif
+
 void main() {
     uint idx = get_idx();
+    uint orig_idx = idx;
 
     uint ne = p.ne20 * p.ne21 * p.ne22 * p.ne23;
 
     // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
     const uint num_iter = 2;
 
+    FLOAT_TYPE sum_sq = 0;
+
     [[unroll]] for (uint i = 0; i < num_iter; ++i) {
         if (idx >= ne) {
             continue;
@@ -61,8 +80,32 @@ void main() {
         [[unroll]] for (uint s = 0; s < num_srcs; ++s) {
             sum += FLOAT_TYPE(a[s].data_a[src_idx(s, i00, i01, i02, i03)]);
         }
+        sum_sq += sum*sum;
         d[num_srcs].data_d[dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
 
         idx += num_threads;
     }
+
+#if ADD_RMS
+    if (p.rms_partials != 0) {
+        // reduce the sum within each subgroup, then across subgroups
+        const uint NumSubgroups = num_threads / gl_SubgroupSize;
+        sum_sq = subgroupAdd(sum_sq);
+        if (gl_SubgroupInvocationID == 0) {
+            sumsh[gl_SubgroupID] = sum_sq;
+        }
+        barrier();
+        [[unroll]] for (uint s = NumSubgroups / 2; s > 0; s >>= 1) {
+            if (gl_SubgroupID < s && gl_SubgroupInvocationID == 0) {
+                sum_sq += sumsh[gl_SubgroupID + s];
+                sumsh[gl_SubgroupID] = sum_sq;
+            }
+            barrier();
+        }
+
+        if (gl_SubgroupID == 0 && gl_SubgroupInvocationID == 0) {
+            partials[num_srcs + 1].partial_sums[orig_idx / (num_iter * num_threads)] = sum_sq;
+        }
+    }
+#endif
 }

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

@@ -10,9 +10,9 @@ layout (constant_id = 1) const bool do_multiply = false;
 
 layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 
-shared FLOAT_TYPE sum[BLOCK_SIZE];
+shared FLOAT_TYPE sumsh[BLOCK_SIZE];
 
-void main() {
+void rms_norm(uint num_iters) {
     const uint ncols     = p.ne00;
     const uint nrows     = gl_NumWorkGroups.x;
     const uint nchannels = gl_NumWorkGroups.y;
@@ -30,38 +30,76 @@ void main() {
     uint32_t b_offset = src1_idx(0, row, channel, samp) + get_boffset();
     uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
 
-    sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
+    FLOAT_TYPE sum = FLOAT_TYPE(0.0f); // partial sum for thread in warp
 
-    [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
-        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[a_offset + col]);
-        sum[tid] += xi * xi;
+    [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
+        FLOAT_TYPE xi = FLOAT_TYPE(0);
+        if (col < ncols) {
+            xi = FLOAT_TYPE(data_a[a_offset + col]);
+        }
+        sum += xi * xi;
     }
 
+    sumsh[tid] = sum;
     // sum up partial sums and write back result
     barrier();
     [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) {
-            sum[tid] += sum[tid + s];
+            sum += sumsh[tid + s];
+            sumsh[tid] = sum;
         }
         barrier();
     }
+    sum = sumsh[0];
 
-    const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(ncols);
+    const FLOAT_TYPE mean = sum / FLOAT_TYPE(ncols);
     const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
 
     if (do_multiply) {
         if (ncols > p.ne10) {
-            [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+            [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
+                if (col >= ncols) {
+                    continue;
+                }
                 data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]));
             }
         } else {
-            [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+            [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
+                if (col >= ncols) {
+                    continue;
+                }
                 data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
             }
         }
     } else {
-        [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+        [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
+            if (col >= ncols) {
+                continue;
+            }
             data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
         }
     }
 }
+
+void main() {
+    // instantiate the rms_norm function for several different
+    // dimensions, to allow loop unrolling
+    uint num_blocks = (p.ne00 + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    if (num_blocks > 32) {
+        rms_norm(num_blocks);
+    } else if (num_blocks > 16) {
+        rms_norm(32);
+    } else if (num_blocks > 8) {
+        rms_norm(16);
+    } else if (num_blocks > 4) {
+        rms_norm(8);
+    } else if (num_blocks == 4) {
+        rms_norm(4);
+    } else if (num_blocks == 3) {
+        rms_norm(3);
+    } else if (num_blocks == 2) {
+        rms_norm(2);
+    } else if (num_blocks == 1) {
+        rms_norm(1);
+    }
+}

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

@@ -0,0 +1,65 @@
+#version 450
+
+#include "generic_binary_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+
+#define BLOCK_SIZE 128
+
+layout (constant_id = 1) const bool do_multiply = false;
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 3, std430) readonly buffer PartialsBuf {float partial_sums[];};
+
+shared FLOAT_TYPE sumsh[BLOCK_SIZE];
+
+void main() {
+    const uint ncols     = p.ne00;
+    const uint nrows     = gl_NumWorkGroups.x;
+    const uint nchannels = gl_NumWorkGroups.y;
+
+    const uint row       = 0;
+    const uint channel   = gl_WorkGroupID.y;
+    const uint samp      = gl_WorkGroupID.z;
+    // The work is split across multiple workgroups in the x dimension. Each invocation
+    // processes one element
+    const uint tid       = gl_GlobalInvocationID.x;
+
+    const uint stride_row       = p.nb01;
+    const uint stride_channel   = p.nb02;
+    const uint stride_sample    = p.nb03;
+
+    uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset();
+    uint32_t b_offset = src1_idx(0, row, channel, samp) + get_boffset();
+    uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
+
+    FLOAT_TYPE sum = FLOAT_TYPE(0.0f); // partial sum for thread in warp
+
+    uint32_t num_partials = p.param3;
+    for (uint32_t i = gl_SubgroupInvocationID; i < num_partials; i += gl_SubgroupSize) {
+        sum += partial_sums[i];
+    }
+    sum = subgroupAdd(sum);
+
+    uint col = tid;
+    if (col >= ncols) {
+        return;
+    }
+
+    const FLOAT_TYPE mean = sum / FLOAT_TYPE(ncols);
+    const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
+
+    if (do_multiply) {
+        if (ncols > p.ne10) {
+            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]));
+        } else {
+            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
+        }
+    } else {
+        data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
+    }
+}

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

@@ -1,9 +1,9 @@
 #version 450
 
-#include "generic_head.comp"
 #include "types.comp"
 
 #extension GL_EXT_control_flow_attributes : enable
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -11,16 +11,49 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 
+layout (push_constant) uniform parameter
+{
+    uint n_cols;
+    uint ne01, ne02;
+    uint nb01, nb02, nb03;
+    uint nb11, nb12, nb13;
+    float weight;
+    uint misalign_offsets;
+    uint ne0_12mp, ne0_12L;
+    uint ne0_1mp, ne0_1L;
+} p;
+
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
+
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
 
 void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint col = gl_LocalInvocationID.x;
+    const float weight = p.weight;
 
-    tmp[col] = FLOAT_TYPE(0.0f);
+    const uint i03 = fastdiv(row, p.ne0_12mp, p.ne0_12L);
+    const uint i03_offset = i03 * p.ne01*p.ne02;
+    const uint i02 = fastdiv(row - i03_offset, p.ne0_1mp, p.ne0_1L);
+    const uint i01 = row - i03_offset - i02*p.ne01;
 
-    for (uint i = col; i < p.KX; i += BLOCK_SIZE) {
-        tmp[col] += FLOAT_TYPE(data_a[row*p.KX + i]);
+    const uint src_idx = get_aoffset() + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+    const uint dst_idx = get_doffset() + i01 * p.nb11 + i02 * p.nb12 + i03 * p.nb13;
+
+    tmp[col] = FLOAT_TYPE(0.0);
+
+    for (uint i = col; i < p.n_cols; i += BLOCK_SIZE) {
+        tmp[col] += FLOAT_TYPE(data_a[src_idx + i]);
     }
 
     barrier();
@@ -32,6 +65,6 @@ void main() {
     }
 
     if (col == 0) {
-        data_d[row] = D_TYPE(tmp[0]);
+        data_d[dst_idx] = D_TYPE(tmp[0] * weight);
     }
 }

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

@@ -68,6 +68,12 @@ const std::vector<std::string> type_names = {
     "bf16",
 };
 
+enum MatMulIdType {
+    NONE,
+    DEFAULT,
+    SUBGROUP,
+};
+
 namespace {
 void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) {
 #ifdef _WIN32
@@ -293,7 +299,7 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const
     compiles.push_back(std::async(string_to_spv_func, _name, in_fname, defines, fp16, coopmat, coopmat2, f16acc));
 }
 
-void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool f16acc) {
+void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool coopmat2, bool f16acc) {
     std::string load_vec = coopmat2 ? "1" : fp16 ? "8" : "4";
     std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4";
     std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
@@ -303,9 +309,13 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
     };
     std::string shader_name = "matmul";
 
-    if (matmul_id) {
+    if (matmul_id_type == MatMulIdType::DEFAULT) {
         base_dict["MUL_MAT_ID"] = "1";
         shader_name = "matmul_id";
+    } else if (matmul_id_type == MatMulIdType::SUBGROUP) {
+        base_dict["MUL_MAT_ID"] = "1";
+        base_dict["MUL_MAT_ID_USE_SUBGROUPS"] = "1";
+        shader_name = "matmul_id_subgroup";
     }
 
     if (fp16) {
@@ -389,7 +399,7 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
         }
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-        if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
+        if (!coopmat && !coopmat2 && matmul_id_type == MatMulIdType::NONE && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
             string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
         }
 #endif
@@ -401,26 +411,28 @@ void process_shaders() {
     std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}};
 
     // matmul
-    for (const auto& matmul_id : {false, true}) {
+    for (const MatMulIdType& matmul_id_type : {MatMulIdType::NONE, MatMulIdType::DEFAULT, MatMulIdType::SUBGROUP}) {
         // No coopmats
         // fp32
-        matmul_shaders(false, matmul_id, false, false, false);
+        matmul_shaders(false, matmul_id_type, false, false, false);
 
         // fp16, fp32acc and fp16acc
-        matmul_shaders(true, matmul_id, false, false, false);
-        matmul_shaders(true, matmul_id, false, false, true);
+        matmul_shaders(true, matmul_id_type, false, false, false);
+        matmul_shaders(true, matmul_id_type, false, false, true);
 
+        if (matmul_id_type != MatMulIdType::DEFAULT) {
 #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
-        // Coopmat, fp32acc and fp16acc
-        matmul_shaders(true, matmul_id, true, false, false);
-        matmul_shaders(true, matmul_id, true, false, true);
+            // Coopmat, fp32acc and fp16acc
+            matmul_shaders(true, matmul_id_type, true, false, false);
+            matmul_shaders(true, matmul_id_type, true, false, true);
 #endif
 
 #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
-        // Coopmat2, fp32acc and fp16acc
-        matmul_shaders(true, matmul_id, false, true, false);
-        matmul_shaders(true, matmul_id, false, true, true);
+            // Coopmat2, fp32acc and fp16acc
+            matmul_shaders(true, matmul_id_type, false, true, false);
+            matmul_shaders(true, matmul_id_type, false, true, true);
 #endif
+        }
     }
 
     // flash attention
@@ -503,6 +515,7 @@ void process_shaders() {
     string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("rms_norm_partials_f32", "rms_norm_partials.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("l2_norm_f32", "l2_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
@@ -538,13 +551,15 @@ void process_shaders() {
         s += std::string(dst_f16 ? "_f16" : "_f32");
         return s;
     };
-    for (std::string op : {"add", "sub", "mul", "div"}) {
+    for (std::string op : {"add", "sub", "mul", "div", "add_rms", }) {
     for (auto src0_f16 : {false, true}) {
     for (auto src1_f16 : {false, true}) {
     for (auto dst_f16  : {false, true}) {
     for (auto rte      : {false, true}) {
+        auto source = op == "add_rms" ? std::string("add") : op;
         auto name = op + get_suffix(src0_f16, src1_f16, dst_f16) + (rte ? "_rte" : "");
-        string_to_spv(name.c_str(), op + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
+        auto add_rms = op == "add_rms" ? "1" : "0";
+        string_to_spv(name.c_str(), source + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}, {"ADD_RMS" , add_rms}});
     }
     }
     }
@@ -586,6 +601,8 @@ void process_shaders() {
 
     string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
 
+    string_to_spv("exp_f16",        "exp.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("exp_f32",        "exp.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("gelu_f16",       "gelu.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("gelu_f32",       "gelu.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("gelu_erf_f16",   "gelu_erf.comp",    {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
@@ -678,12 +695,15 @@ void process_shaders() {
 
     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"}}));
+    string_to_spv("conv2d_dw_whcn_f16_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
+    string_to_spv("conv2d_dw_cwhn_f16_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
 
     string_to_spv("roll_f32", "roll.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     string_to_spv("add_id_f32", "add_id.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
 
-    string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
+    string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}, {"ADD_RMS" , "0"}});
+    string_to_spv("multi_add_rms_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}, {"ADD_RMS" , "1"}});
 
     for (auto &c : compiles) {
         c.wait();
@@ -741,7 +761,7 @@ void write_output_files() {
     }
 
     std::string suffixes[2] = {"_f32", "_f16"};
-    for (const char *op : {"add", "sub", "mul", "div"}) {
+    for (const char *op : {"add", "sub", "mul", "div", "add_rms"}) {
         fprintf(hdr, "extern unsigned char *%s_data[2][2][2][2];\n", op);
         fprintf(hdr, "extern uint64_t %s_len[2][2][2][2];\n", op);
         std::string data = "unsigned char *" + std::string(op) + "_data[2][2][2][2] = ";

ggml/src/ggml-webgpu/CMakeLists.txt

@@ -20,8 +20,8 @@ add_custom_command(
     COMMAND ${CMAKE_COMMAND} -E make_directory ${SHADER_OUTPUT_DIR}
     COMMAND ${CMAKE_COMMAND} -E env PYTHONIOENCODING=utf-8
         ${Python3_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders/embed_wgsl.py
-            --input "${SHADER_DIR}"
-            --output "${SHADER_HEADER}"
+            --input_dir "${SHADER_DIR}"
+            --output_file "${SHADER_HEADER}"
     DEPENDS ${WGSL_SHADER_FILES} ${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders/embed_wgsl.py
     VERBATIM
 )

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -118,13 +118,11 @@ struct webgpu_context_struct {
 
     std::recursive_mutex mutex;
 
-    bool device_init = false;
-
     webgpu_buf_pool param_buf_pool;
     webgpu_buf_pool set_rows_error_buf_pool;
 
     wgpu::ComputePipeline memset_pipeline;
-    wgpu::ComputePipeline mul_mat_pipeline;
+    wgpu::ComputePipeline mul_mat_pipeline[30][2];
     wgpu::ComputePipeline set_rows_pipeline;
     wgpu::ComputePipeline cpy_pipeline;
 
@@ -238,7 +236,7 @@ static void ggml_backend_webgpu_wait_on_submission(webgpu_context & ctx) {
                                   wgpu::CallbackMode::AllowSpontaneous,
                                   [](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
                                       if (status != wgpu::QueueWorkDoneStatus::Success) {
-                                          GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", message.data);
+                                          GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", std::string(message).c_str());
                                       }
                                   }),
                               UINT64_MAX);
@@ -278,7 +276,7 @@ static void ggml_backend_webgpu_submit_queue(webgpu_context & ctx) {
         wgpu::CallbackMode::AllowSpontaneous,
         [ctx, staged_param_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
             if (status != wgpu::QueueWorkDoneStatus::Success) {
-                GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", message.data);
+                GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", std::string(message).c_str());
             }
             // Free the staged buffers
             ctx->param_buf_pool.free_bufs(staged_param_bufs);
@@ -294,7 +292,7 @@ static void ggml_backend_webgpu_submit_queue(webgpu_context & ctx) {
             wgpu::CallbackMode::AllowSpontaneous,
             [ctx, error_bufs](wgpu::MapAsyncStatus status, wgpu::StringView message) {
                 if (status != wgpu::MapAsyncStatus::Success) {
-                    GGML_LOG_ERROR("ggml_webgpu: Failed to map error buffer: %s\n", message.data);
+                    GGML_LOG_ERROR("ggml_webgpu: Failed to map error buffer: %s\n", std::string(message).c_str());
                 } else {
                     const uint32_t * error_data = (const uint32_t *) error_bufs.host_buf.GetConstMappedRange();
                     if (*error_data) {
@@ -331,6 +329,7 @@ static void ggml_backend_webgpu_map_buffer(webgpu_context & ctx,
 // To use, add a bind group entry to the setup for the shader you are debugging, add the buffer and
 // debug statements in the shader, and then call this function after encoding the commands and submitting them.
 static void ggml_backend_webgpu_debug(webgpu_context & ctx) {
+    ggml_backend_webgpu_submit_queue(ctx);
     wgpu::CommandEncoder encoder = ctx->device.CreateCommandEncoder();
     encoder.CopyBufferToBuffer(ctx->debug_dev_buf, 0, ctx->debug_host_buf, 0, ctx->debug_host_buf.GetSize());
     wgpu::CommandBuffer commands = encoder.Finish();
@@ -421,15 +420,6 @@ static void ggml_backend_webgpu_buffer_memset(webgpu_context & ctx,
     ggml_backend_webgpu_build_and_enqueue(ctx, ctx->memset_pipeline, params, entries, wg_x, true);
 }
 
-static size_t ggml_backend_webgpu_tensor_offset(const ggml_tensor * tensor) {
-    return webgpu_tensor_offset(tensor) + tensor->view_offs;
-}
-
-static wgpu::Buffer ggml_backend_webgpu_tensor_buf(const ggml_tensor * tensor) {
-    ggml_backend_webgpu_buffer_context * ctx = (ggml_backend_webgpu_buffer_context *) tensor->buffer->context;
-    return ctx->buffer;
-}
-
 /** End WebGPU Actions */
 
 /** GGML Backend Interface */
@@ -447,19 +437,36 @@ static void ggml_backend_webgpu_free(ggml_backend_t backend) {
     GGML_UNUSED(ctx);
 }
 
+static size_t ggml_webgpu_tensor_offset(const ggml_tensor * tensor) {
+    return webgpu_tensor_offset(tensor) + tensor->view_offs;
+}
+
+static wgpu::Buffer ggml_webgpu_tensor_buf(const ggml_tensor * tensor) {
+    ggml_backend_webgpu_buffer_context * ctx = (ggml_backend_webgpu_buffer_context *) tensor->buffer->context;
+    return ctx->buffer;
+}
+
+static size_t ggml_webgpu_tensor_misalignment(webgpu_context & ctx, ggml_tensor * t) {
+    size_t offset = ggml_webgpu_tensor_offset(t);
+    return offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
+}
+
+static size_t ggml_webgpu_tensor_align_offset(webgpu_context & ctx, ggml_tensor * t) {
+    size_t offset = ggml_webgpu_tensor_offset(t);
+    return offset & ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
+}
+
+static size_t ggml_webgpu_tensor_binding_size(webgpu_context & ctx, ggml_tensor * t) {
+    return (ggml_nbytes(t) + ggml_webgpu_tensor_misalignment(ctx, t) + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) &
+           ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1);
+}
+
 static void ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
-    size_t src_offset       = ggml_backend_webgpu_tensor_offset(src);
-    // assumes power of 2 offset alignment
-    size_t src_misalignment = src_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
-    // align to minimum offset alignment
-    src_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
-    size_t dst_offset       = ggml_backend_webgpu_tensor_offset(dst);
-    size_t dst_misalignment = dst_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
-    dst_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
-    uint32_t              ne     = (uint32_t) ggml_nelements(dst);
+    uint32_t ne = (uint32_t) ggml_nelements(dst);
+
     std::vector<uint32_t> params = { ne,
-                                     (uint32_t) (src_misalignment / ggml_type_size(src->type)),
-                                     (uint32_t) (dst_misalignment / ggml_type_size(dst->type)),
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src) / ggml_type_size(src->type)),
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
                                      // Convert byte-strides to element-strides
                                      (uint32_t) (src->nb[0] / ggml_type_size(src->type)),
                                      (uint32_t) (src->nb[1] / ggml_type_size(src->type)),
@@ -477,15 +484,13 @@ static void ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, ggml_tensor
 
     std::vector<wgpu::BindGroupEntry> entries = {
         { .binding = 0,
-         .buffer  = ggml_backend_webgpu_tensor_buf(src),
-         .offset  = src_offset,
-         .size    = (ggml_nbytes(src) + src_misalignment + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) &
-                  ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1) },
+         .buffer  = ggml_webgpu_tensor_buf(src),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src) },
         { .binding = 1,
-         .buffer  = ggml_backend_webgpu_tensor_buf(dst),
-         .offset  = dst_offset,
-         .size    = (ggml_nbytes(dst) + dst_misalignment + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) &
-                  ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1) }
+         .buffer  = ggml_webgpu_tensor_buf(dst),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, dst) }
     };
 
     size_t   max_wg_size = ctx->limits.maxComputeWorkgroupSizeX;
@@ -504,21 +509,9 @@ static void ggml_webgpu_set_rows(webgpu_context & ctx, ggml_tensor * src, ggml_t
         error_bufs.host_buf.Unmap();
     }
 
-    size_t src_offset       = ggml_backend_webgpu_tensor_offset(src);
-    // assumes power of 2 offset alignment
-    size_t src_misalignment = src_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
-    // align to minimum offset alignment
-    src_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
-    size_t idx_offset       = ggml_backend_webgpu_tensor_offset(idx);
-    size_t idx_misalignment = idx_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
-    idx_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
-    size_t dst_offset       = ggml_backend_webgpu_tensor_offset(dst);
-    size_t dst_misalignment = dst_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
-    dst_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
-
-    std::vector<uint32_t> params = { (uint32_t) (src_misalignment / ggml_type_size(src->type)),
-                                     (uint32_t) (idx_misalignment / ggml_type_size(idx->type)),
-                                     (uint32_t) (dst_misalignment / ggml_type_size(dst->type)),
+    std::vector<uint32_t> params = { (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src) / ggml_type_size(src->type)),
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, idx) / ggml_type_size(idx->type)),
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
                                      // Convert byte-strides to element-strides
                                      (uint32_t) (src->nb[1] / ggml_type_size(src->type)),
                                      (uint32_t) (src->nb[2] / ggml_type_size(src->type)),
@@ -540,18 +533,18 @@ static void ggml_webgpu_set_rows(webgpu_context & ctx, ggml_tensor * src, ggml_t
 
     std::vector<wgpu::BindGroupEntry> entries = {
         { .binding = 0,
-         .buffer  = ggml_backend_webgpu_tensor_buf(src),
-         .offset  = ggml_backend_webgpu_tensor_offset(src),
-         .size    = ggml_nbytes(src)                                                                       },
+         .buffer  = ggml_webgpu_tensor_buf(src),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src) },
         { .binding = 1,
-         .buffer  = ggml_backend_webgpu_tensor_buf(idx),
-         .offset  = ggml_backend_webgpu_tensor_offset(idx),
-         .size    = ggml_nbytes(idx)                                                                       },
+         .buffer  = ggml_webgpu_tensor_buf(idx),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, idx),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, idx) },
         { .binding = 2,
-         .buffer  = ggml_backend_webgpu_tensor_buf(dst),
-         .offset  = ggml_backend_webgpu_tensor_offset(dst),
-         .size    = ggml_nbytes(dst)                                                                       },
-        { .binding = 3, .buffer = error_bufs.dev_buf,    .offset = 0, .size = error_bufs.dev_buf.GetSize() }
+         .buffer  = ggml_webgpu_tensor_buf(dst),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, dst) },
+        { .binding = 3, .buffer = error_bufs.dev_buf, .offset = 0, .size = error_bufs.dev_buf.GetSize() }
     };
 
     size_t   max_wg_size = ctx->limits.maxComputeWorkgroupSizeX;
@@ -565,15 +558,18 @@ static void ggml_webgpu_set_rows(webgpu_context & ctx, ggml_tensor * src, ggml_t
 
 static void ggml_webgpu_mul_mat(webgpu_context & ctx, ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst) {
     std::vector<uint32_t> params = {
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
         (uint32_t) dst->ne[1],                                  // number of rows in result (M)
         (uint32_t) dst->ne[0],                                  // number of columns in result (N)
         (uint32_t) src0->ne[0],                                 // number of columns in src0/src1 (K)
-        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),  // stride (elements) of src0 in dimension 1
-        (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),  // stride (elements) of src1 in dimension 1
-        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),  // stride (elements) of src0 in dimension 2
-        (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),  // stride (elements) of src1 in dimension 2
-        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),  // stride (elements) of src0 in dimension 3
-        (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),  // stride (elements) of src1 in dimension 3
+        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),  // stride (elements/blocks) of src0 in dimension 1
+        (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),  // stride (elements/blocks) of src1 in dimension 1
+        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),  // stride (elements/blocks) of src0 in dimension 2
+        (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),  // stride (elements/blocks) of src1 in dimension 2
+        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),  // stride (elements/blocks) of src0 in dimension 3
+        (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),  // stride (elements/blocks) of src1 in dimension 3
         (uint32_t) src0->ne[2],                                 // batch size in dimension 2
         (uint32_t) src0->ne[3],                                 // batch size in dimension 3
         (uint32_t) (src1->ne[2] / src0->ne[2]),                 // broadcast in dimension 2
@@ -582,22 +578,22 @@ static void ggml_webgpu_mul_mat(webgpu_context & ctx, ggml_tensor * src0, ggml_t
 
     std::vector<wgpu::BindGroupEntry> entries = {
         { .binding = 0,
-         .buffer  = ggml_backend_webgpu_tensor_buf(src0),
-         .offset  = ggml_backend_webgpu_tensor_offset(src0),
-         .size    = ggml_nbytes(src0) },
+         .buffer  = ggml_webgpu_tensor_buf(src0),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src0),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src0) },
         { .binding = 1,
-         .buffer  = ggml_backend_webgpu_tensor_buf(src1),
-         .offset  = ggml_backend_webgpu_tensor_offset(src1),
-         .size    = ggml_nbytes(src1) },
+         .buffer  = ggml_webgpu_tensor_buf(src1),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src1),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src1) },
         { .binding = 2,
-         .buffer  = ggml_backend_webgpu_tensor_buf(dst),
-         .offset  = ggml_backend_webgpu_tensor_offset(dst),
-         .size    = ggml_nbytes(dst)  }
+         .buffer  = ggml_webgpu_tensor_buf(dst),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, dst)  },
     };
 
     uint32_t wg_x =
         (dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3] + WEBGPU_MUL_MAT_WG_SIZE - 1) / WEBGPU_MUL_MAT_WG_SIZE;
-    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->mul_mat_pipeline, params, entries, wg_x);
+    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->mul_mat_pipeline[src0->type][src1->type], params, entries, wg_x);
 }
 
 // Returns true if node has enqueued work into the queue, false otherwise
@@ -827,7 +823,7 @@ static ggml_backend_buffer_t ggml_backend_webgpu_buffer_type_alloc_buffer(ggml_b
     wgpu::Buffer buf;
     ggml_webgpu_create_buffer(ctx->webgpu_ctx->device,
                               buf,
-                              size,
+                              (size + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) & ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1),
                               wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::CopyDst,
                               "allocated_buffer");
 
@@ -907,7 +903,94 @@ static void ggml_webgpu_init_memset_pipeline(webgpu_context & webgpu_ctx) {
 }
 
 static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) {
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline, wgsl_mul_mat, "mul_mat");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F32][GGML_TYPE_F32],
+                                wgsl_mul_mat_f32_f32,
+                                "mul_mat_f32_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F16][GGML_TYPE_F16],
+                                wgsl_mul_mat_f16_f16,
+                                "mul_mat_f16_f16");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F16][GGML_TYPE_F32],
+                                wgsl_mul_mat_f16_f32,
+                                "mul_mat_f16_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_0][GGML_TYPE_F32],
+                                wgsl_mul_mat_q4_0_f32,
+                                "mul_mat_q4_0_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_1][GGML_TYPE_F32],
+                                wgsl_mul_mat_q4_1_f32,
+                                "mul_mat_q4_1_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q5_0][GGML_TYPE_F32],
+                                wgsl_mul_mat_q5_0_f32,
+                                "mul_mat_q5_0_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q5_1][GGML_TYPE_F32],
+                                wgsl_mul_mat_q5_1_f32,
+                                "mul_mat_q5_1_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q8_0][GGML_TYPE_F32],
+                                wgsl_mul_mat_q8_0_f32,
+                                "mul_mat_q8_0_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q2_K][GGML_TYPE_F32],
+                                wgsl_mul_mat_q2_k_f32,
+                                "mul_mat_q2_k_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q3_K][GGML_TYPE_F32],
+                                wgsl_mul_mat_q3_k_f32,
+                                "mul_mat_q3_k_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_K][GGML_TYPE_F32],
+                                wgsl_mul_mat_q4_k_f32,
+                                "mul_mat_q4_k_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q5_K][GGML_TYPE_F32],
+                                wgsl_mul_mat_q5_k_f32,
+                                "mul_mat_q5_k_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q6_K][GGML_TYPE_F32],
+                                wgsl_mul_mat_q6_k_f32,
+                                "mul_mat_q6_k_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ2_XXS][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq2_xxs_f32,
+                                "mul_mat_iq2_xxs_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ2_XS][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq2_xs_f32,
+                                "mul_mat_iq2_xs_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ2_S][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq2_s_f32,
+                                "mul_mat_iq2_s_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ3_XXS][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq3_xxs_f32,
+                                "mul_mat_iq3_xxs_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ3_S][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq3_s_f32,
+                                "mul_mat_iq3_s_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ1_S][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq1_s_f32,
+                                "mul_mat_iq1_s_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ1_M][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq1_m_f32,
+                                "mul_mat_iq1_m_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ4_NL][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq4_nl_f32,
+                                "mul_mat_iq4_nl_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ4_XS][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq4_xs_f32,
+                                "mul_mat_iq4_xs_f32");
 }
 
 static void ggml_webgpu_init_set_rows_pipeline(webgpu_context & webgpu_ctx) {
@@ -933,79 +1016,6 @@ static ggml_backend_t ggml_backend_webgpu_device_init(ggml_backend_dev_t dev, co
     ggml_backend_webgpu_device_context * dev_ctx    = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
     webgpu_context                       webgpu_ctx = dev_ctx->webgpu_ctx;
 
-    // Multiple threads may try to initialize the device
-    std::lock_guard<std::recursive_mutex> lock(webgpu_ctx->mutex);
-    if (!webgpu_ctx->device_init) {
-        // Initialize device
-        std::vector<wgpu::FeatureName> required_features = { wgpu::FeatureName::ShaderF16,
-                                                             wgpu::FeatureName::ImplicitDeviceSynchronization };
-        wgpu::DeviceDescriptor         dev_desc;
-        dev_desc.requiredLimits       = &webgpu_ctx->limits;
-        dev_desc.requiredFeatures     = required_features.data();
-        dev_desc.requiredFeatureCount = required_features.size();
-        dev_desc.SetDeviceLostCallback(
-            wgpu::CallbackMode::AllowSpontaneous,
-            [](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) {
-                GGML_UNUSED(device);
-                GGML_LOG_ERROR(
-                    "ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason), message.data);
-            });
-        dev_desc.SetUncapturedErrorCallback(
-            [](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) {
-                GGML_UNUSED(device);
-                GGML_LOG_ERROR(
-                    "ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason), message.data);
-            });
-        webgpu_ctx->instance.WaitAny(
-            webgpu_ctx->adapter.RequestDevice(
-                &dev_desc,
-                wgpu::CallbackMode::AllowSpontaneous,
-                [webgpu_ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
-                    if (status != wgpu::RequestDeviceStatus::Success) {
-                        GGML_LOG_ERROR("ggml_webgpu: Failed to get a device: %s\n", message.data);
-                        return;
-                    }
-                    webgpu_ctx->device = std::move(device);
-                }),
-            UINT64_MAX);
-        GGML_ASSERT(webgpu_ctx->device != nullptr);
-
-        // Initialize (compute) queue
-        webgpu_ctx->queue = webgpu_ctx->device.GetQueue();
-
-        // Create buffer pool for shader parameters
-        webgpu_ctx->param_buf_pool.init(webgpu_ctx->device,
-                                        WEBGPU_NUM_PARAM_BUFS,
-                                        WEBGPU_PARAMS_BUF_SIZE_BYTES,
-                                        wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform,
-                                        wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite);
-        webgpu_ctx->set_rows_error_buf_pool.init(webgpu_ctx->device,
-                                                 WEBGPU_NUM_SET_ROWS_ERROR_BUFS,
-                                                 WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES,
-                                                 wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::Storage,
-                                                 wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead);
-
-        ggml_webgpu_init_memset_pipeline(webgpu_ctx);
-        ggml_webgpu_init_mul_mat_pipeline(webgpu_ctx);
-        ggml_webgpu_init_set_rows_pipeline(webgpu_ctx);
-        ggml_webgpu_init_cpy_pipeline(webgpu_ctx);
-
-#ifdef GGML_WEBGPU_DEBUG
-        // Initialize debug buffers
-        ggml_webgpu_create_buffer(webgpu_ctx->device,
-                                  webgpu_ctx->debug_host_buf,
-                                  WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
-                                  wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead,
-                                  "debug_host_buf");
-        ggml_webgpu_create_buffer(webgpu_ctx->device,
-                                  webgpu_ctx->debug_dev_buf,
-                                  WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
-                                  wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc,
-                                  "debug_dev_buf");
-#endif
-        webgpu_ctx->device_init = true;
-    }
-
     static ggml_backend_webgpu_context backend_ctx;
     backend_ctx.name       = GGML_WEBGPU_NAME + std::string(": ") + dev_ctx->device_name;
     backend_ctx.webgpu_ctx = webgpu_ctx;
@@ -1053,10 +1063,45 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
             return true;
-        case GGML_OP_CPY | GGML_OP_SET_ROWS:
+        case GGML_OP_CPY:
+        case GGML_OP_SET_ROWS:
             return op->type == GGML_TYPE_F16 && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_MUL_MAT:
-            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
+            {
+                switch (op->src[1]->type) {
+                    case GGML_TYPE_F16:
+                        return op->src[0]->type == GGML_TYPE_F16;
+                    case GGML_TYPE_F32:
+                        switch (op->src[0]->type) {
+                            case GGML_TYPE_F32:
+                            case GGML_TYPE_F16:
+                            case GGML_TYPE_Q4_0:
+                            case GGML_TYPE_Q4_1:
+                            case GGML_TYPE_Q5_0:
+                            case GGML_TYPE_Q5_1:
+                            case GGML_TYPE_Q8_0:
+                            case GGML_TYPE_Q2_K:
+                            case GGML_TYPE_Q3_K:
+                            case GGML_TYPE_Q4_K:
+                            case GGML_TYPE_Q5_K:
+                            case GGML_TYPE_Q6_K:
+                            case GGML_TYPE_IQ2_XXS:
+                            case GGML_TYPE_IQ2_XS:
+                            case GGML_TYPE_IQ2_S:
+                            case GGML_TYPE_IQ3_XXS:
+                            case GGML_TYPE_IQ3_S:
+                            case GGML_TYPE_IQ1_S:
+                            case GGML_TYPE_IQ1_M:
+                            case GGML_TYPE_IQ4_NL:
+                            case GGML_TYPE_IQ4_XS:
+                                return true;
+                            default:
+                                return false;
+                        }
+                    default:
+                        return false;
+                }
+            }
         default:
             return false;
     }
@@ -1123,20 +1168,87 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
     wgpu::AdapterInfo info{};
     ctx->adapter.GetInfo(&info);
 
+    // Initialize device
+    std::vector<wgpu::FeatureName> required_features = { wgpu::FeatureName::ShaderF16,
+                                                         wgpu::FeatureName::ImplicitDeviceSynchronization };
+    wgpu::DeviceDescriptor         dev_desc;
+    dev_desc.requiredLimits       = &ctx->limits;
+    dev_desc.requiredFeatures     = required_features.data();
+    dev_desc.requiredFeatureCount = required_features.size();
+    dev_desc.SetDeviceLostCallback(
+        wgpu::CallbackMode::AllowSpontaneous,
+        [](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) {
+            GGML_UNUSED(device);
+            GGML_LOG_ERROR(
+                "ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason), std::string(message).c_str());
+        });
+    dev_desc.SetUncapturedErrorCallback(
+        [](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) {
+            GGML_UNUSED(device);
+            GGML_LOG_ERROR(
+                "ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason), std::string(message).c_str());
+        });
+    ctx->instance.WaitAny(ctx->adapter.RequestDevice(
+                              &dev_desc,
+                              wgpu::CallbackMode::AllowSpontaneous,
+                              [ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
+                                  if (status != wgpu::RequestDeviceStatus::Success) {
+                                      GGML_LOG_ERROR("ggml_webgpu: Failed to get a device: %s\n", std::string(message).c_str());
+                                      return;
+                                  }
+                                  ctx->device = std::move(device);
+                              }),
+                          UINT64_MAX);
+    GGML_ASSERT(ctx->device != nullptr);
+
+    // Initialize (compute) queue
+    ctx->queue = ctx->device.GetQueue();
+
+    // Create buffer pool for shader parameters
+    ctx->param_buf_pool.init(ctx->device,
+                             WEBGPU_NUM_PARAM_BUFS,
+                             WEBGPU_PARAMS_BUF_SIZE_BYTES,
+                             wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform,
+                             wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite);
+    ctx->set_rows_error_buf_pool.init(ctx->device,
+                                      WEBGPU_NUM_SET_ROWS_ERROR_BUFS,
+                                      WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES,
+                                      wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::Storage,
+                                      wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead);
+
+    ggml_webgpu_init_memset_pipeline(ctx);
+    ggml_webgpu_init_mul_mat_pipeline(ctx);
+    ggml_webgpu_init_set_rows_pipeline(ctx);
+    ggml_webgpu_init_cpy_pipeline(ctx);
+
+#ifdef GGML_WEBGPU_DEBUG
+    // Initialize debug buffers
+    ggml_webgpu_create_buffer(ctx->device,
+                              ctx->debug_host_buf,
+                              WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
+                              wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead,
+                              "debug_host_buf");
+    ggml_webgpu_create_buffer(ctx->device,
+                              ctx->debug_dev_buf,
+                              WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
+                              wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc,
+                              "debug_dev_buf");
+#endif
+
     static ggml_backend_webgpu_device_context device_ctx;
     device_ctx.webgpu_ctx  = ctx;
     device_ctx.device_name = GGML_WEBGPU_NAME;
-    device_ctx.device_desc = std::string(info.description.data);
+    device_ctx.device_desc = info.description;
 
     GGML_LOG_INFO(
         "ggml_webgpu: adapter_info: vendor_id: %u | vendor: %s | architecture: %s | device_id: %u | name: %s | "
         "device_desc: %s\n",
         info.vendorID,
-        info.vendor.data,
-        info.architecture.data,
+        std::string(info.vendor).c_str(),
+        std::string(info.architecture).c_str(),
         info.deviceID,
-        info.device.data,
-        info.description.data);
+        std::string(info.device).c_str(),
+        std::string(info.description).c_str());
 
     // See GGML Backend Device Interface section
     static ggml_backend_device device = {

ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py

@@ -1,35 +1,85 @@
 import os
+import re
+import ast
 import argparse
 
 
-def escape_triple_quotes(wgsl):
-    # Simple defense in case of embedded """
-    return wgsl.replace('"""', '\\"""')
+def extract_block(text, name):
+    pattern = rf'#define\({name}\)\s*(.*?)#end\({name}\)'
+    match = re.search(pattern, text, re.DOTALL)
+    if not match:
+        raise ValueError(f"Missing block: {name}")
+    return match.group(1).strip()
 
 
-def to_cpp_string_literal(varname, content):
-    return f'const char* wgsl_{varname} = R"({content})";\n'
+def parse_decls(decls_text):
+    decls = {}
+    for name, code in re.findall(r'#decl\((.*?)\)\s*(.*?)#enddecl\(\1\)', decls_text, re.DOTALL):
+        decls[name.strip()] = code.strip()
+    return decls
+
+
+def replace_placeholders(shader_text, replacements):
+    for key, val in replacements.items():
+        # Match {{KEY}} literally, where KEY is escaped
+        pattern = r'{{\s*' + re.escape(key) + r'\s*}}'
+        shader_text = re.sub(pattern, str(val), shader_text)
+    return shader_text
+
+
+def write_shader(shader_name, shader_code, output_dir, outfile):
+    if output_dir:
+        wgsl_filename = os.path.join(output_dir, f"{shader_name}.wgsl")
+        with open(wgsl_filename, "w", encoding="utf-8") as f_out:
+            f_out.write(shader_code)
+    outfile.write(f'const char* wgsl_{shader_name} = R"({shader_code})";\n\n')
+
+
+def generate_variants(shader_path, output_dir, outfile):
+    shader_base_name = shader_path.split("/")[-1].split(".")[0]
+
+    with open(shader_path, "r", encoding="utf-8") as f:
+        text = f.read()
+
+    try:
+        variants = ast.literal_eval(extract_block(text, "VARIANTS"))
+    except ValueError:
+        write_shader(shader_base_name, text, output_dir, outfile)
+    else:
+        decls_map = parse_decls(extract_block(text, "DECLS"))
+        shader_template = extract_block(text, "SHADER")
+
+        for variant in variants:
+            decls = variant["DECLS"]
+            decls_code = ""
+            for key in decls:
+                if key not in decls_map:
+                    raise ValueError(f"DECLS key '{key}' not found.")
+                decls_code += decls_map[key] + "\n\n"
+
+            shader_variant = replace_placeholders(shader_template, variant["REPLS"])
+            final_shader = re.sub(r'\bDECLS\b', decls_code, shader_variant)
+
+            output_name = f"{shader_base_name}_" + "_".join([variant["REPLS"]["SRC0_TYPE"], variant["REPLS"]["SRC1_TYPE"]])
+            write_shader(output_name, final_shader, output_dir, outfile)
 
 
 def main():
     parser = argparse.ArgumentParser()
-    parser.add_argument('--input', required=True)
-    parser.add_argument('--output', required=True)
+    parser.add_argument("--input_dir", required=True)
+    parser.add_argument("--output_file", required=True)
+    parser.add_argument("--output_dir")
     args = parser.parse_args()
 
-    with open(args.output, 'w', encoding='utf-8') as out:
-        out.write("// Auto-generated shader embedding \n\n")
-        for fname in sorted(os.listdir(args.input)):
-            if not fname.endswith('.wgsl'):
-                continue
-            shader_path = os.path.join(args.input, fname)
-            varname = os.path.splitext(fname)[0]
-            with open(shader_path, 'r', encoding='utf-8') as f:
-                content = f.read()
-            content = escape_triple_quotes(content)
-            out.write(to_cpp_string_literal(varname, content))
-            out.write('\n')
+    if args.output_dir:
+        os.makedirs(args.output_dir, exist_ok=True)
+
+    with open(args.output_file, "w", encoding="utf-8") as out:
+        out.write("// Auto-generated shader embedding\n\n")
+        for fname in sorted(os.listdir(args.input_dir)):
+            if fname.endswith(".wgsl"):
+                generate_variants(os.path.join(args.input_dir, fname), args.output_dir, out)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl

@@ -19,20 +19,20 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let start = params.offset;
     let end = params.offset + params.size;
 
-    for (var j: u32 = 0u; j < bytes_per_thread; j = j + 1u) {
+    for (var j: u32 = 0u; j < bytes_per_thread; j += 4) {
         let byte_index = start + i + j;
-        if (byte_index + 4u <= end) {
-            output_buffer[(byte_index >> 2u)] = params.value;
+        if (byte_index + 4 <= end) {
+            output_buffer[byte_index >> 2] = params.value;
         } else {
             // Handle tail (unaligned)
-            for (var k: u32 = 0u; k < 4u; k = k + 1u) {
+            for (var k: u32 = 0; k < 4; k++) {
                 let idx = byte_index + k;
                 if (idx < end) {
-                    let word_idx = idx >> 2u;
-                    let byte_offset = (idx & 3u) * 8u;
-                    let mask = ~(0xffu << byte_offset);
+                    let word_idx = idx >> 2;
+                    let bit_offset = (idx & 3) * 8u;
+                    let mask = ~(0xffu << bit_offset);
                     let existing = output_buffer[word_idx];
-                    output_buffer[word_idx] = (existing & mask) | ((params.value & 0xffu) << byte_offset);
+                    output_buffer[word_idx] = (existing & mask) | (params.value & (0xffu << bit_offset));
                 }
             }
         }

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl

@@ -1,56 +0,0 @@
-struct MulMatParams {
-    m: u32,
-    n: u32,
-    k: u32,
-    // all strides are in elements
-    stride_01: u32,
-    stride_11: u32,
-    stride_02: u32,
-    stride_12: u32,
-    stride_03: u32,
-    stride_13: u32,
-
-    bs02: u32,
-    bs03: u32,
-    broadcast2: u32,
-    broadcast3: u32
-};
-
-@group(0) @binding(0) var<storage, read_write> src0: array<f32>; // N rows, K columns
-@group(0) @binding(1) var<storage, read_write> src1: array<f32>; // M rows, K columns (transposed)
-@group(0) @binding(2) var<storage, read_write> dst: array<f32>; // M rows, N columns
-
-@group(0) @binding(3) var<uniform> params: MulMatParams;
-
-@compute @workgroup_size(64)
-fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
-    let total = params.m * params.n * params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
-    if (global_id.x >= total) {
-        return;
-    }
-
-    let dst2_stride = params.m * params.n;
-    let dst3_stride = dst2_stride * params.bs02 * params.broadcast2;
-
-    let dst3_idx = global_id.x / dst3_stride;
-    let src03_idx = dst3_idx / params.broadcast3; // src0 may be broadcast along the third dimension
-    let src13_idx = dst3_idx; // src1 is not broadcast
-    let dst3_rem = global_id.x % dst3_stride;
-
-    let dst2_idx = dst3_rem / dst2_stride;
-    let src02_idx = dst2_idx / params.broadcast2; // src0 may also be broadcast along the second dimension
-    let src12_idx = dst2_idx; // src1 is not broadcast
-
-    let dst2_rem = dst3_rem % dst2_stride;
-
-    let row = dst2_rem / params.n; // output row
-    let col = dst2_rem % params.n; // output column
-
-    var sum = 0.0;
-    for (var i: u32 = 0u; i < params.k; i = i + 1u) {
-        let src0_idx = src03_idx * params.stride_03 + src02_idx * params.stride_02 + col * params.stride_01 + i;
-        let src1_idx = src13_idx * params.stride_13 + src12_idx * params.stride_12 + row * params.stride_11 + i;
-        sum = sum + src0[src0_idx] * src1[src1_idx];
-    }
-    dst[dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.n + col] = sum;
-}

ggml/src/ggml.c

@@ -975,6 +975,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "IM2COL",
     "IM2COL_BACK",
     "CONV_2D",
+    "CONV_3D",
     "CONV_2D_DW",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
@@ -1017,7 +1018,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 88, "GGML_OP_COUNT != 88");
+static_assert(GGML_OP_COUNT == 89, "GGML_OP_COUNT != 89");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1077,6 +1078,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "im2col(x)",
     "im2col_back(x)",
     "conv_2d(x)",
+    "conv_3d(x)",
     "conv_2d_dw(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
@@ -1119,7 +1121,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 88, "GGML_OP_COUNT != 88");
+static_assert(GGML_OP_COUNT == 89, "GGML_OP_COUNT != 89");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -4480,6 +4482,56 @@ struct ggml_tensor * ggml_conv_2d_direct(
     return result;
 }
 
+// ggml_conv_3d
+
+struct ggml_tensor * ggml_conv_3d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int                   s0,
+        int                   s1,
+        int                   s2,
+        int                   p0,
+        int                   p1,
+        int                   p2,
+        int                   d0,
+        int                   d1,
+        int                   d2,
+        int                   c,
+        int                   n,
+        int                   oc) {
+
+    GGML_ASSERT(a->ne[3] == (int64_t) c * oc);
+    GGML_ASSERT(b->ne[3] == (int64_t) c * n);
+
+    int64_t ne[4];
+    ne[0] = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+    ne[1] = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1);
+    ne[2] = ggml_calc_conv_output_size(b->ne[2], a->ne[2], s2, p2, d2);
+    ne[3] = (int64_t) oc * n;
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+
+    ggml_set_op_params_i32(result, 0,  s0);
+    ggml_set_op_params_i32(result, 1,  s1);
+    ggml_set_op_params_i32(result, 2,  s2);
+    ggml_set_op_params_i32(result, 3,  p0);
+    ggml_set_op_params_i32(result, 4,  p1);
+    ggml_set_op_params_i32(result, 5,  p2);
+    ggml_set_op_params_i32(result, 6,  d0);
+    ggml_set_op_params_i32(result, 7,  d1);
+    ggml_set_op_params_i32(result, 8,  d2);
+    ggml_set_op_params_i32(result, 9,  c);
+    ggml_set_op_params_i32(result, 10, n);
+    ggml_set_op_params_i32(result, 11, oc);
+
+    result->op = GGML_OP_CONV_3D;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
 // ggml_conv_transpose_2d_p0
 
 static int64_t ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int s, int p) {

gguf-py/gguf/constants.py

@@ -385,6 +385,7 @@ class MODEL_ARCH(IntEnum):
     DREAM            = auto()
     SMALLTHINKER     = auto()
     LLADA            = auto()
+    SEED_OSS         = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -717,6 +718,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.DREAM:            "dream",
     MODEL_ARCH.SMALLTHINKER:     "smallthinker",
     MODEL_ARCH.LLADA:            "llada",
+    MODEL_ARCH.SEED_OSS:         "seed_oss",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1973,6 +1975,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.SEED_OSS: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+    ],
     MODEL_ARCH.OLMOE: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -2590,6 +2606,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ATTN_K,
         MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.OUTPUT,
     ],
     MODEL_ARCH.SMALLTHINKER: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -2833,6 +2850,7 @@ class VisionProjectorType:
     QWEN25O = "qwen2.5o" # omni
     VOXTRAL = "voxtral"
     LFM2 = "lfm2"
+    KIMIVL = "kimivl"
 
 
 # Items here are (block size, type size)