cuda: fix build warnings in set-rows.cu (unused variable) (#14687)

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

.github/workflows/update-ops-docs.yml

@@ -0,0 +1,40 @@
+name: Update Operations Documentation
+
+on:
+    push:
+        paths:
+            - 'docs/ops/**'
+            - 'scripts/create_ops_docs.py'
+    pull_request:
+        paths:
+            - 'docs/ops/**'
+            - 'scripts/create_ops_docs.py'
+
+jobs:
+    update-ops-docs:
+        runs-on: ubuntu-latest
+
+        steps:
+        - name: Checkout repository
+          uses: actions/checkout@v4
+
+        - name: Set up Python
+          uses: actions/setup-python@v5
+          with:
+              python-version: '3.x'
+
+        - name: Generate operations documentation to temporary file
+          run: |
+              mkdir -p /tmp/ops_check
+              ./scripts/create_ops_docs.py /tmp/ops_check/ops.md
+
+        - name: Check if docs/ops.md matches generated version
+          run: |
+              if ! diff -q docs/ops.md /tmp/ops_check/ops.md; then
+                  echo "Operations documentation (docs/ops.md) is not up to date with the backend CSV files."
+                  echo "To fix: run ./scripts/create_ops_docs.py and commit the updated docs/ops.md along with your changes"
+                  echo "Differences found:"
+                  diff docs/ops.md /tmp/ops_check/ops.md || true
+                  exit 1
+              fi
+              echo "Operations documentation is up to date."

CMakePresets.json

@@ -55,6 +55,17 @@
             "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-apple-clang.cmake"
         }
     },
+    {
+        "name": "x64-linux-gcc", "hidden": true,
+        "cacheVariables": {
+            "CMAKE_C_COMPILER": "gcc",
+            "CMAKE_CXX_COMPILER": "g++"
+        }
+    },
+    { "name": "x64-linux-gcc-debug", "inherits": [ "base", "x64-linux-gcc", "debug" ] },
+    { "name": "x64-linux-gcc-release", "inherits": [ "base", "x64-linux-gcc", "release" ] },
+    { "name": "x64-linux-gcc-reldbg", "inherits": [ "base", "x64-linux-gcc", "reldbg" ] },
+    { "name": "x64-linux-gcc+static-release", "inherits": [ "base", "x64-linux-gcc", "release", "static" ] },
 
     { "name": "arm64-windows-llvm-debug", "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
     { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },

README.md

@@ -6,9 +6,9 @@
 [![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 
-[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
+[Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md)
 
-Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
+LLM inference in C/C++
 
 ## Recent API changes
 
@@ -17,10 +17,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
-- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
+- Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
+- Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
-- Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
 - Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
@@ -134,6 +133,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
 - [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
+- [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
 
 #### Multimodal
 

common/CMakeLists.txt

@@ -86,8 +86,7 @@ if (LLAMA_CURL)
     endif()
     target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
     include_directories(${CURL_INCLUDE_DIRS})
-    find_library(CURL_LIBRARY curl REQUIRED)
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARIES})
 endif ()
 
 if (LLAMA_LLGUIDANCE)
@@ -112,8 +111,8 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.7.20 (+ fix to build on GCC 15):
-        GIT_TAG b5b8b64dba11c4e4ee6b1d1450d3a3ae279891e8
+        # v1.0.1:
+        GIT_TAG d795912fedc7d393de740177ea9ea761e7905774
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE

convert_hf_to_gguf.py

@@ -300,6 +300,7 @@ class ModelBase:
                             gguf.MODEL_TENSOR.POS_EMBD,
                             gguf.MODEL_TENSOR.TOKEN_TYPES,
                             gguf.MODEL_TENSOR.SSM_CONV1D,
+                            gguf.MODEL_TENSOR.SHORTCONV_CONV,
                             gguf.MODEL_TENSOR.TIME_MIX_FIRST,
                             gguf.MODEL_TENSOR.TIME_MIX_W1,
                             gguf.MODEL_TENSOR.TIME_MIX_W2,
@@ -833,6 +834,12 @@ class TextModel(ModelBase):
         if chkhsh == "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b":
             # ref: https://huggingface.co/tiiuae/Falcon-H1-34B-Base
             res = "falcon-h1"
+        if chkhsh == "f6791d196f87ce6b56a7d234be618e0d58f8cda3549416635b2bebcd22cd95c4":
+            # ref: https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct
+            res = "midm-2.0"
+        if chkhsh == "169bf0296a13c4d9b7672313f749eb36501d931022de052aad6e36f2bf34dd51":
+            # ref: https://huggingface.co/LiquidAI/LFM2-Tokenizer
+            res = "lfm2"
 
         if res is None:
             logger.warning("\n")
@@ -1075,7 +1082,14 @@ class TextModel(ModelBase):
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
-        special_vocab.chat_template = "rwkv-world"
+        if special_vocab.chat_template is None:
+            template_path = Path(__file__).parent / "models" / "templates" / "llama-cpp-rwkv-world.jinja"
+            if template_path.is_file():
+                with open(template_path, "r", encoding="utf-8") as f:
+                    template = f.read()
+            else:
+                template = "rwkv-world"
+            special_vocab.chat_template = template
         # hack: Add '\n\n' as the EOT token to make it chat normally
         special_vocab._set_special_token("eot", 261)
         # hack: Override these as they have already been set (incorrectly)
@@ -4890,6 +4904,9 @@ class Mamba2Model(TextModel):
             with open(dir_model / "config.json", "r", encoding="utf-8") as f:
                 hparams = json.load(f)
         super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+        self.d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
+        self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * self.d_model
+        self.n_group = self.find_hparam(["n_groups"], optional=True) or 1
 
     def set_vocab(self):
         vocab_size = self.hparams["vocab_size"]
@@ -4912,12 +4929,9 @@ class Mamba2Model(TextModel):
             self._set_vocab_builtin("gpt-neox", vocab_size)
 
     def set_gguf_parameters(self):
-        d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
-        d_conv  = self.find_hparam(["conv_kernel",       "d_conv"],  optional=True) or 4
-        d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * d_model
-        d_state = self.find_hparam(["state_size",        "d_state"], optional=True) or 128
-        head_dim = self.find_hparam(["mamba_d_head", "head_dim"],    optional=True) or 64
-        n_group = self.find_hparam(["n_groups"],                     optional=True) or 1
+        d_conv  = self.find_hparam(["conv_kernel", "d_conv"],     optional=True) or 4
+        d_state = self.find_hparam(["state_size",  "d_state"],    optional=True) or 128
+        head_dim = self.find_hparam(["mamba_d_head", "head_dim"], optional=True) or 64
 
         rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
 
@@ -4925,19 +4939,19 @@ class Mamba2Model(TextModel):
         # TODO: does this really matter?
         # skip the assertion for FalconH1 Model
         if self.model_arch != gguf.MODEL_ARCH.FALCON_H1:
-            assert d_inner == 2 * d_model
-            assert d_inner % head_dim == 0
+            assert self.d_inner == 2 * self.d_model
+            assert self.d_inner % head_dim == 0
 
         self.gguf_writer.add_context_length(2**20)  # arbitrary value; for those who use the default
-        self.gguf_writer.add_embedding_length(d_model)
+        self.gguf_writer.add_embedding_length(self.d_model)
         self.gguf_writer.add_feed_forward_length(0)  # unused, but seemingly required when loading
         self.gguf_writer.add_head_count(0)  # unused, but seemingly required when loading
         self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_ssm_conv_kernel(d_conv)
-        self.gguf_writer.add_ssm_inner_size(d_inner)
+        self.gguf_writer.add_ssm_inner_size(self.d_inner)
         self.gguf_writer.add_ssm_state_size(d_state)
-        self.gguf_writer.add_ssm_time_step_rank(d_inner // head_dim)
-        self.gguf_writer.add_ssm_group_count(n_group)
+        self.gguf_writer.add_ssm_time_step_rank(self.d_inner // head_dim)
+        self.gguf_writer.add_ssm_group_count(self.n_group)
         self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
         self.gguf_writer.add_file_type(self.ftype)
 
@@ -4962,10 +4976,7 @@ class Mamba2Model(TextModel):
             # (D is also unsqueezed, but for more straightforward broadcast internally)
             data_torch = data_torch.reshape((*data_torch.shape, 1))
         elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_NORM, bid):
-            d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
-            d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * d_model
-            n_group = self.hparams.get("n_groups", 1)
-            data_torch = data_torch.reshape((n_group, d_inner // n_group))
+            data_torch = data_torch.reshape((self.n_group, self.d_inner // self.n_group))
 
         if name.endswith(".A_log"):
             logger.debug("A_log --> A ==> " + new_name)
@@ -6452,18 +6463,148 @@ class GraniteMoeModel(GraniteModel):
                 (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
             ]
 
+        has_experts = bool(self.hparams.get('num_local_experts'))
+
         if name.endswith("shared_mlp.input_linear.weight"):
             ffn_dim = self.hparams["shared_intermediate_size"]
             assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
             gate, up = data_torch.split(ffn_dim, dim=-2)
+            if has_experts:
+                return [
+                    (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
+                    (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+                ]
             return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), gate),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), up),
+            ]
+
+        if not has_experts and name.endswith("shared_mlp.output_linear.weight"):
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), data_torch)
             ]
 
         return super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("GraniteMoeHybridForCausalLM", "BambaForCausalLM")
+class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
+    """GraniteHybrid is a hybrid SSM + Attention model that uses Mamba2 SSM
+    layers and optionally uses MoE w/ a shared expert"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_HYBRID
+    undo_permute = True
+
+    def __init__(self, *args, **kwargs):
+
+        # Hybrid mamba models use a prefix for the mamba-specific params.
+        # TODO: Extend this if the prefix(es) need to be configurable
+        self.hparam_prefixes = ["mamba"]
+
+        super().__init__(*args, **kwargs)
+
+        # Lists of which layers use ssm vs attention
+        self._attn_layers = self.get_attn_layers()
+        self._ssm_layers = [
+            i for i in range(self.block_count)
+            if i not in self._attn_layers
+        ]
+
+        # n_group and d_inner are used during reshape_tensors for mamba2
+        self.d_model = self.find_hparam(["hidden_size", "d_model"])
+        self.n_group = self.find_hparam(["n_groups"])
+        self.d_inner = self.find_hparam(["expand"]) * self.d_model
+
+    def get_attn_layers(self):
+        # Explicit list of layer type names
+        if layer_types := self.hparams.get("layer_types"):
+            return [
+                i for i, typ in enumerate(layer_types)
+                if typ == "attention"
+            ]
+
+        # Layer types indicated by index or period
+        attn_layers = self.hparams.get("attn_layer_indices", [])
+        if not attn_layers:
+            attn_period = self.hparams.get("attn_layer_period")
+            assert attn_period, "Didn't find attn_layer_indices or attn_layer_period"
+            attn_offset = self.hparams.get("attn_layer_offset")
+            assert attn_offset is not None, "No attention layer offset set with attn_layer_period"
+            attn_layers = [
+                i for i in range(self.block_count)
+                if i % attn_period == attn_offset
+            ]
+        return attn_layers
+
+    def find_hparam(self, keys: Iterable[str], *args, **kwargs) -> Any:
+        prefixed = []
+        for pfx in self.hparam_prefixes:
+            prefixed.extend(
+                "_".join([pfx, k])
+                for k in keys
+            )
+        keys = list(keys) + prefixed
+        return Mamba2Model.find_hparam(self, keys, *args, **kwargs)
+
+    def modify_tensors(
+        self, data_torch: Tensor, name: str, bid: int | None
+    ) -> Iterable[tuple[str, Tensor]]:
+        if (
+            name.endswith("block_sparse_moe.input_linear.weight")
+            or "shared_mlp" in name
+        ):
+            return GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
+
+        # Determine whether this is a mamba layer or an attention layer
+        if bid in self._ssm_layers:
+            return Mamba2Model.modify_tensors(self, data_torch, name, bid)
+        elif bid in self._attn_layers:
+            return GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def set_gguf_parameters(self):
+        """This method merges params from both parents and some that are
+        specific to this model. The result is some duplication of how the params
+        get set. The following warnings are expected during conversion:
+
+        WARNING:Duplicated key name 'granitehybrid.attention.head_count_kv'
+        WARNING:Duplicated key name 'granitehybrid.context_length'
+        """
+        GraniteMoeModel.set_gguf_parameters(self)
+
+        ## Mamba mixer params ##
+        self.gguf_writer.add_ssm_conv_kernel(self.find_hparam(["conv_kernel", "d_conv"]))
+        self.gguf_writer.add_ssm_state_size(self.find_hparam(["state_size", "d_state"]))
+        self.gguf_writer.add_ssm_group_count(self.n_group)
+        self.gguf_writer.add_ssm_inner_size(self.d_inner)
+        # NOTE: The mamba_dt_rank is _not_ the right field for how this is used
+        #   in llama.cpp
+        self.gguf_writer.add_ssm_time_step_rank(self.find_hparam(["n_heads"]))
+
+        ## Attention params ##
+        head_count_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
+        head_count_kv_vec = [
+            head_count_kv if i in self._attn_layers else 0 for i in range(self.block_count)
+        ]
+        if rope_dim := self.hparams.get("attn_rotary_emb"):
+            self.gguf_writer.add_rope_dimension_count(rope_dim)
+        self.gguf_writer.add_head_count_kv(head_count_kv_vec)
+
+        ## If Bamba, use rope, otherwise don't
+        use_rope = "BambaForCausalLM" in self.hparams["architectures"]
+        self.gguf_writer.add_rope_scaling_finetuned(use_rope)
+        if not use_rope:
+            self.gguf_writer.add_context_length(2**20)
+
+        ## Validation ##
+        d_head = self.find_hparam(["d_head"], optional=True) or 64
+        assert self.hparams.get("hidden_act") in [None, "silu"], "Only SILU activation supported"
+        assert self.d_inner % d_head == 0, f"SSM inner size {self.d_inner} not a multiple of head dim {d_head}"
+
+    def set_vocab(self):
+        self.hparams["pad_vocab_size_multiple"] = 8
+        Mamba2Model.set_vocab(self)
+
+
 @ModelBase.register("BailingMoeForCausalLM")
 class BailingMoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.BAILINGMOE
@@ -6687,7 +6828,7 @@ class FalconH1Model(Mamba2Model):
         # Use Llama conversion for attention
         self._transformer_model_class = LlamaModel
 
-        # n_group and d_inner are used during reshape_tensors for mamaba2
+        # n_group and d_inner are used during reshape_tensors for mamba2
         self.n_group = self.find_hparam(["n_groups"])
         self.d_inner = self.find_hparam(["mamba_d_ssm"])
         self.d_head = self.find_hparam(["d_head"])
@@ -6943,6 +7084,50 @@ class SmolLM3Model(LlamaModel):
             chat_template = tokenizer.chat_template.replace("[:]", "")
             self.gguf_writer.add_chat_template(chat_template)
 
+
+@ModelBase.register("Lfm2ForCausalLM")
+@ModelBase.register("LFM2ForCausalLM")
+class LFM2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.LFM2
+
+    def _add_feed_forward_length(self):
+        ff_dim = self.hparams["block_ff_dim"]
+
+        auto_adjust_ff_dim = self.hparams["block_auto_adjust_ff_dim"]
+        ff_dim = self.hparams["block_ff_dim"]
+        ffn_dim_multiplier = self.hparams["block_ffn_dim_multiplier"]
+        multiple_of = self.hparams["block_multiple_of"]
+
+        if auto_adjust_ff_dim:
+            ff_dim = int(2 * ff_dim / 3)
+            # custom dim factor multiplier
+            if ffn_dim_multiplier is not None:
+                ff_dim = int(ffn_dim_multiplier * ff_dim)
+            ff_dim = multiple_of * ((ff_dim + multiple_of - 1) // multiple_of)
+
+        self.gguf_writer.add_feed_forward_length(ff_dim)
+
+    def set_gguf_parameters(self):
+        # set num_key_value_heads only for attention layers
+        self.hparams["num_key_value_heads"] = [
+            self.hparams["num_key_value_heads"] if layer_type == "full_attention" else 0
+            for layer_type in self.hparams["layer_types"]
+        ]
+
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+        self.gguf_writer.add_shortconv_l_cache(self.hparams["conv_L_cache"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["norm_eps"])
+        self._add_feed_forward_length()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # conv op requires 2d tensor
+        if 'conv.conv' in name:
+            data_torch = data_torch.squeeze(1)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 ###### CONVERSION LOGIC ######
 
 

convert_hf_to_gguf_update.py

@@ -129,6 +129,8 @@ models = [
     {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
     {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
     {"name": "a.x-4.0",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/skt/A.X-4.0", },
+    {"name": "midm-2.0",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },
+    {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions

docs/ops.md

@@ -0,0 +1,95 @@
+# GGML Operations
+
+List of GGML operations and backend support status.
+
+Legend:
+- ✅ Fully supported by this backend
+- 🟡 Partially supported by this backend
+- ❌ Not supported by this backend
+
+| Operation | BLAS | CPU | CUDA | Metal |
+|-----------|------|------|------|------|
+|                              ABS | ❌ | ✅ | 🟡 | ❌ |
+|                              ACC | ❌ | ✅ | ✅ | ✅ |
+|                              ADD | ❌ | ✅ | ✅ | 🟡 |
+|                             ADD1 | ❌ | ✅ | ✅ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ |
+|                           ARGMAX | ❌ | ✅ | ✅ | ✅ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ |
+|                            CLAMP | ❌ | ✅ | ✅ | 🟡 |
+|                           CONCAT | ❌ | ✅ | 🟡 | ✅ |
+|                             CONT | ❌ | ✅ | 🟡 | ✅ |
+|                       CONV_2D_DW | ❌ | ✅ | ✅ | ❌ |
+|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ |
+|                CONV_TRANSPOSE_2D | ❌ | ✅ | ✅ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | 🟡 |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ❌ |
+|                              CPY | ❌ | 🟡 | 🟡 | 🟡 |
+|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ✅ | ✅ | ❌ |
+|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | 🟡 |
+|                              DIV | ❌ | ✅ | ✅ | 🟡 |
+|                              DUP | ❌ | ✅ | 🟡 | 🟡 |
+|                              ELU | ❌ | ✅ | ❌ | 🟡 |
+|                              EXP | ❌ | ✅ | 🟡 | ❌ |
+|                   FLASH_ATTN_EXT | ❌ | ✅ | 🟡 | 🟡 |
+|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ❌ |
+|                            GEGLU | ❌ | ✅ | ✅ | 🟡 |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | 🟡 |
+|                      GEGLU_QUICK | ❌ | ✅ | ✅ | 🟡 |
+|                             GELU | ❌ | ✅ | 🟡 | 🟡 |
+|                         GELU_ERF | ❌ | ✅ | 🟡 | 🟡 |
+|                       GELU_QUICK | ❌ | ✅ | 🟡 | 🟡 |
+|                         GET_ROWS | ❌ | ✅ | 🟡 | ✅ |
+|                    GET_ROWS_BACK | ❌ | 🟡 | 🟡 | ❌ |
+|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ |
+|                      HARDSIGMOID | ❌ | ✅ | 🟡 | ❌ |
+|                        HARDSWISH | ❌ | ✅ | 🟡 | ❌ |
+|                           IM2COL | ❌ | ✅ | ✅ | 🟡 |
+|                          L2_NORM | ❌ | ✅ | ✅ | ✅ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ |
+|                              LOG | ❌ | ✅ | ✅ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ |
+|                              MUL | ❌ | ✅ | ✅ | 🟡 |
+|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 |
+|                       MUL_MAT_ID | ❌ | ✅ | ✅ | ✅ |
+|                              NEG | ❌ | ✅ | 🟡 | 🟡 |
+|                             NORM | ❌ | ✅ | ✅ | 🟡 |
+|                   OPT_STEP_ADAMW | ❌ | ✅ | ✅ | ❌ |
+|                         OUT_PROD | 🟡 | 🟡 | 🟡 | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | ✅ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ❌ | ✅ |
+|                          POOL_2D | ❌ | ✅ | ✅ | ✅ |
+|                            REGLU | ❌ | ✅ | ✅ | 🟡 |
+|                             RELU | ❌ | ✅ | 🟡 | 🟡 |
+|                           REPEAT | ❌ | ✅ | 🟡 | ✅ |
+|                      REPEAT_BACK | ❌ | ✅ | ✅ | ❌ |
+|                         RMS_NORM | ❌ | ✅ | ✅ | 🟡 |
+|                    RMS_NORM_BACK | ❌ | ✅ | ✅ | ❌ |
+|                     RMS_NORM_MUL | ❌ | ✅ | ✅ | ✅ |
+|                             ROPE | ❌ | ✅ | ✅ | ✅ |
+|                        ROPE_BACK | ❌ | ✅ | ✅ | ❌ |
+|                        RWKV_WKV6 | ❌ | ✅ | ✅ | ✅ |
+|                        RWKV_WKV7 | ❌ | ✅ | ✅ | ✅ |
+|                            SCALE | ❌ | ✅ | ✅ | ✅ |
+|                              SET | ❌ | ✅ | ❌ | ✅ |
+|                         SET_ROWS | ❌ | 🟡 | ❌ | 🟡 |
+|                              SGN | ❌ | ✅ | 🟡 | ❌ |
+|                          SIGMOID | ❌ | ✅ | 🟡 | 🟡 |
+|                             SILU | ❌ | ✅ | 🟡 | 🟡 |
+|                        SILU_BACK | ❌ | ✅ | ✅ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | 🟡 |
+|                         SOFT_MAX | ❌ | ✅ | ✅ | ✅ |
+|                    SOFT_MAX_BACK | ❌ | 🟡 | 🟡 | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | 🟡 |
+|                             SQRT | ❌ | ✅ | ✅ | 🟡 |
+|                         SSM_CONV | ❌ | ✅ | ✅ | ✅ |
+|                         SSM_SCAN | ❌ | ✅ | ✅ | ✅ |
+|                             STEP | ❌ | ✅ | 🟡 | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | 🟡 |
+|                              SUM | ❌ | ✅ | ✅ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ |
+|                           SWIGLU | ❌ | ✅ | ✅ | 🟡 |
+|                             TANH | ❌ | ✅ | 🟡 | 🟡 |
+|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ |
+|                          UPSCALE | ❌ | ✅ | ✅ | 🟡 |

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

@@ -2090,6 +2090,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             {
                 // TODO: add support
                 // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+#pragma message("TODO: implement F32, F16, BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
                 return false;
             } break;
         case GGML_OP_CPY: {

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

@@ -43,6 +43,7 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/set-rows.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -2230,6 +2231,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_GET_ROWS_BACK:
             ggml_cuda_op_get_rows_back(ctx, dst);
             break;
+        case GGML_OP_SET_ROWS:
+            ggml_cuda_op_set_rows(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cuda_dup(ctx, dst);
             break;
@@ -2299,6 +2303,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_EXP:
                     ggml_cuda_op_exp(ctx, dst);
                     break;
+                case GGML_UNARY_OP_ELU:
+                    ggml_cuda_op_elu(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -3112,6 +3119,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_ELU:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -3216,6 +3224,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->ne[2] == 1 && op->ne[3] == 1;
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+#pragma message("TODO: implement Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
+                return (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16) &&
+                       op->src[0]->type == GGML_TYPE_F32 &&
+                       op->src[1]->type == GGML_TYPE_I64;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;

ggml/src/ggml-cuda/set-rows.cu

@@ -0,0 +1,151 @@
+#include "set-rows.cuh"
+
+typedef void (*set_rows_kernel_t)(const char * src, char * dst);
+
+template<typename src_t, typename dst_t>
+__device__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
+    GGML_UNUSED(src_f);
+    GGML_UNUSED(dst_f);
+}
+
+template<>
+__device__ __forceinline__ void set_rows_1<float, half>(const float * src_f, half * dst_h) {
+    *dst_h = __float2half(*src_f);
+}
+
+template<>
+__device__ __forceinline__ void set_rows_1<float, nv_bfloat16>(const float * src_f, nv_bfloat16 * dst_b) {
+    *dst_b = *src_f;
+}
+
+template<>
+__device__ __forceinline__ void set_rows_1<float, float>(const float * src_f, float * dst_f) {
+    *dst_f = *src_f;
+}
+
+template<typename src_t, typename dst_t>
+static __global__ void k_set_rows(
+        const src_t * __restrict__ src0, const int64_t * __restrict__ src1, dst_t * __restrict__ dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t s10, const int64_t s11, const int64_t s12,
+        const int64_t s1, const int64_t s2, const int64_t s3) {
+
+    const int64_t i = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
+    const int64_t ne_total = ne00 * ne01 * ne02 * ne03;
+
+    if (i >= ne_total) {
+        return;
+    }
+
+    const int64_t i03 = i / (ne00 * ne01 * ne02);
+    const int64_t i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int64_t i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
+    const int64_t i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+
+    const int64_t i12 = i03 % ne12;
+    const int64_t i11 = i02 % ne11;
+    const int64_t i10 = i01;
+
+    const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
+
+    const src_t * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
+    dst_t * dst_row_ptr    = dst + dst_row*s1 + i02*s2 + i03*s3;
+
+    const src_t* src_elem = src0_row + i00;
+    dst_t* dst_elem = dst_row_ptr + i00;
+    set_rows_1(src_elem, dst_elem);
+
+    GGML_UNUSED(ne10);
+    GGML_UNUSED(ne13);
+}
+
+template<typename src_t, typename dst_t>
+static void set_rows_cuda(
+        const src_t * src0_d, const int64_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+
+    const int64_t ne_total = ne00 * ne01 * ne02 * ne03;
+    const int num_blocks = (ne_total + CUDA_SET_ROWS_BLOCK_SIZE - 1) / CUDA_SET_ROWS_BLOCK_SIZE;
+    const dim3 block_size(CUDA_SET_ROWS_BLOCK_SIZE);
+    const dim3 grid_size(num_blocks);
+
+
+    const int64_t s01 = nb01/sizeof(src_t);
+    const int64_t s02 = nb02/sizeof(src_t);
+    const int64_t s03 = nb03/sizeof(src_t);
+    const int64_t s10 = nb10/sizeof(int64_t);
+    const int64_t s11 = nb11/sizeof(int64_t);
+    const int64_t s12 = nb12/sizeof(int64_t);
+    const int64_t s1  = nb1/sizeof(dst_t);
+    const int64_t s2  = nb2/sizeof(dst_t);
+    const int64_t s3  = nb3/sizeof(dst_t);
+
+    if (ne_total > 0) {
+        k_set_rows<<<grid_size, block_size, 0, stream>>>(
+            src0_d, src1_d, dst_d,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            s01, s02, s03,
+            s10, s11, s12,
+            s1, s2, s3);
+    }
+}
+
+
+void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_I64);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const float * src0_d   = (const float *)src0->data;
+    const int64_t * src1_d = (const int64_t *)src1->data;
+
+    cudaStream_t stream = ctx.stream();
+
+
+
+    if (dst->type == GGML_TYPE_F32) {
+        set_rows_cuda(
+            src0_d, src1_d, (float*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_F16) {
+        set_rows_cuda(
+            src0_d, src1_d, (half*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_BF16) {
+        set_rows_cuda(
+            src0_d, src1_d, (nv_bfloat16*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else {
+        GGML_ABORT("unsupported type");
+    }
+}

ggml/src/ggml-cuda/set-rows.cuh

@@ -0,0 +1,7 @@
+#pragma once
+
+#include "common.cuh"
+
+#define CUDA_SET_ROWS_BLOCK_SIZE 256
+
+void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/ssm-conv.cu

@@ -107,8 +107,11 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
         if (nc == 4) {
             ssm_conv_f32<threads, 4><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
                                                                      dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+        } else if (nc == 3) {
+            ssm_conv_f32<threads, 3><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
+                                                                     dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
-            GGML_ABORT("Only support kernel size = 4  now.");
+            GGML_ABORT("Only support kernel size = 3 or size = 4 right now.");
         }
     } else {
         if (nc == 4) {
@@ -116,8 +119,13 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
             dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
             ssm_conv_long_token_f32<threads, 4, split_n_t><<<blocks, threads, 0, stream>>>(
                 src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+        } else if (nc == 3) {
+            const int64_t split_n_t = 32;
+            dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
+            ssm_conv_long_token_f32<threads, 3, split_n_t><<<blocks, threads, 0, stream>>>(
+                src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
-            GGML_ABORT("Only support kernel size = 4 right now.");
+            GGML_ABORT("Only support kernel size = 3 or size = 4 right now.");
         }
     }
 }

ggml/src/ggml-cuda/unary.cu

@@ -83,6 +83,10 @@ static __device__ __forceinline__ float op_log(float x) {
     return logf(x);
 }
 
+static __device__ __forceinline__ float op_elu(float x) {
+    return (x > 0.f) ? x : expm1f(x);
+}
+
 template <float (*op)(float), typename T>
 static __global__ void unary_op_kernel(const T * x, T * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
@@ -196,6 +200,9 @@ void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_unary<op_log>(ctx, dst);
 }
 
+void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_elu>(ctx, dst);
+}
 /* gated ops */
 
 template <float (*op)(float), typename T>

ggml/src/ggml-cuda/unary.cuh

@@ -59,6 +59,8 @@ void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -10,9 +10,6 @@
 #include "rocblas/rocblas.h"
 #endif // __HIP_PLATFORM_AMD__
 
-#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
-#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
-#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_OP_N HIPBLAS_OP_N
@@ -30,7 +27,6 @@
 #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_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
-#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
 #define cublasCreate hipblasCreate
 #define cublasDestroy hipblasDestroy
 #define cublasGemmEx hipblasGemmEx
@@ -42,7 +38,6 @@
 #define cublasSgemm hipblasSgemm
 #define cublasStatus_t hipblasStatus_t
 #define cublasOperation_t hipblasOperation_t
-#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
 #define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
 #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
 #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
@@ -144,6 +139,20 @@
 #define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
 #define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
 
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION >= 70000000
+#define CUBLAS_COMPUTE_16F HIPBLAS_COMPUTE_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_COMPUTE_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_COMPUTE_32F_FAST_16F
+#define cublasComputeType_t hipblasComputeType_t
+#define cudaDataType_t hipDataType
+#else
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define cublasComputeType_t hipblasDatatype_t
+#define cudaDataType_t hipblasDatatype_t
+#endif
+
 #define __CUDA_ARCH__ 1300
 
 #if defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__)

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

@@ -173,6 +173,12 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_SILU,
     GGML_METAL_KERNEL_TYPE_SILU_4,
     GGML_METAL_KERNEL_TYPE_ELU,
+    GGML_METAL_KERNEL_TYPE_ABS,
+    GGML_METAL_KERNEL_TYPE_SGN,
+    GGML_METAL_KERNEL_TYPE_STEP,
+    GGML_METAL_KERNEL_TYPE_HARDSWISH,
+    GGML_METAL_KERNEL_TYPE_HARDSIGMOID,
+    GGML_METAL_KERNEL_TYPE_EXP,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,
@@ -1155,6 +1161,12 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                            silu,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4,                          silu_4,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ELU,                             elu,                             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ABS,                             abs,                             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SGN,                             sgn,                             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_STEP,                            step,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_HARDSWISH,                       hardswish,                       true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_HARDSIGMOID,                     hardsigmoid,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_EXP,                             exp,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,                    soft_max_f16,                    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,                  soft_max_f16_4,                  has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,                    soft_max_f32,                    has_simdgroup_reduction);
@@ -1688,6 +1700,12 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_ELU:
                 case GGML_UNARY_OP_NEG:
+                case GGML_UNARY_OP_ABS:
+                case GGML_UNARY_OP_SGN:
+                case GGML_UNARY_OP_STEP:
+                case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_EXP:
                     return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 default:
                     return false;
@@ -2439,6 +2457,78 @@ static bool ggml_metal_encode_node(
 
                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                 } break;
+                case GGML_UNARY_OP_ABS:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ABS].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_SGN:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SGN].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_STEP:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_STEP].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_HARDSWISH:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_HARDSWISH].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_HARDSIGMOID:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_HARDSIGMOID].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_EXP:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_EXP].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
                 default:
                 {
                     GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));

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

@@ -1199,6 +1199,51 @@ kernel void kernel_neg(
     dst[tpig] = -src0[tpig];
 }
 
+kernel void kernel_abs(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = fabs(src0[tpig]);
+}
+
+kernel void kernel_sgn(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = (x > 0.0f) ? 1.0f : ((x < 0.0f) ? -1.0f : 0.0f);
+}
+
+kernel void kernel_step(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] > 0.0f ? 1.0f : 0.0f;
+}
+
+kernel void kernel_hardswish(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = x * fmin(1.0f, fmax(0.0f, (x + 3.0f) / 6.0f));
+}
+
+kernel void kernel_hardsigmoid(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = fmin(1.0f, fmax(0.0f, (x + 3.0f) / 6.0f));
+}
+
+kernel void kernel_exp(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = exp(src0[tpig]);
+}
+
 kernel void kernel_reglu(
         device const char * src0,
         device const char * src1,

ggml/src/ggml-opencl/CMakeLists.txt

@@ -88,6 +88,7 @@ set(GGML_OPENCL_KERNELS
     rms_norm
     rope
     scale
+    set_rows
     sigmoid
     silu
     softmax_4_f32
@@ -103,6 +104,7 @@ set(GGML_OPENCL_KERNELS
     tanh
     pad
     repeat
+    mul_mat_f16_f32
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

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

@@ -351,6 +351,7 @@ struct ggml_backend_opencl_context {
     cl_program program_gemv_noshuffle_general;
     cl_program program_gemv_noshuffle;
     cl_program program_get_rows;
+    cl_program program_set_rows;
     cl_program program_glu;
     cl_program program_im2col_f16;
     cl_program program_im2col_f32;
@@ -367,6 +368,7 @@ struct ggml_backend_opencl_context {
     cl_program program_mul_mv_f16_f32;
     cl_program program_mul_mv_f32_f32;
     cl_program program_mul;
+    cl_program program_mul_mat_f16_f32_tiled;
     cl_program program_div;
     cl_program program_sub;
     cl_program program_norm;
@@ -412,6 +414,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_soft_max, kernel_soft_max_4;
     cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
     cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
+    cl_kernel kernel_set_rows_f32, kernel_set_rows_f16;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
     cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
     cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
@@ -420,6 +423,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul_mat_f16_f32_1row;
     cl_kernel kernel_mul_mat_f16_f32;
     cl_kernel kernel_mul_mat_f16_f32_l4;
+    cl_kernel kernel_mul_mat_f16_f32_tiled;
     cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
     cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
     cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
@@ -529,6 +533,16 @@ struct ggml_backend_opencl_context {
         fclose(ftrace);
     }
 
+    size_t get_kernel_workgroup_size(cl_kernel kernel) const {
+        size_t workgroup_size = 0;
+        size_t ret_size = 0;
+        CL_CHECK(
+            clGetKernelWorkGroupInfo(kernel, device, CL_KERNEL_WORK_GROUP_SIZE,
+                sizeof(size_t), &workgroup_size, &ret_size));
+        GGML_ASSERT(sizeof(size_t) == ret_size);
+        return workgroup_size;
+    }
+
     void enqueue_ndrange_kernel(cl_kernel kernel, cl_uint work_dim, size_t *global_work_size, size_t *local_work_size, const ggml_tensor * tensor) {
 #ifdef GGML_OPENCL_PROFILING
         cl_event evt;
@@ -1003,6 +1017,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mat_f16_f32_tiled
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mat_f16_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mat_f16_f32.cl");
+#endif
+        backend_ctx->program_mul_mat_f16_f32_tiled =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_f16_f32_tiled = clCreateKernel(backend_ctx->program_mul_mat_f16_f32_tiled, "mul_mat_f16_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1431,6 +1461,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         }
     }
 
+    // set_rows
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "set_rows.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("set_rows.cl");
+#endif
+        backend_ctx->program_set_rows =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_set_rows_f32  = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f16  = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_id_q4_0_f32_8x_flat
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2233,8 +2280,18 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             {
                 // TODO: add support
                 // ref: https://github.com/ggml-org/llama.cpp/pull/14274
-                return false;
-            } break;
+#pragma message("TODO: implement BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
+                if (op->src[0]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+                switch (op->type) {
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_F32:
+                        return true;
+                    default:
+                        return false;
+                }
+            }
         case GGML_OP_CPY:
         case GGML_OP_DUP:
         case GGML_OP_CONT:
@@ -3374,6 +3431,111 @@ static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, c
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
+static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    // ne0 = ne00
+    // ne2 = ne02
+    // ne3 = ne03
+
+    const int      ne01 = src0->ne[1];
+    const int      ne02 = src0->ne[2];
+    const int      ne03 = src0->ne[3];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int      ne11 = src1->ne[1];
+    const int      ne12 = src1->ne[2];
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+
+    const int      ne0  = dst->ne[0];
+
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    const int nblk0 = ne0/ggml_blck_size(dst->type);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    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 * 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 offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            kernel = backend_ctx->kernel_set_rows_f32;
+            break;
+        case GGML_TYPE_F16:
+            kernel = backend_ctx->kernel_set_rows_f16;
+            break;
+        default:
+            GGML_ABORT("not implemented");
+    }
+
+    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),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &nblk0));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb3));
+
+    int nth0 = 64;
+    if (backend_ctx->gpu_family == INTEL) {
+        nth0 = 32;
+    } else if (backend_ctx->gpu_family == ADRENO) {
+        nth0 = 64;
+    }
+
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    while (nth0 < nblk0 && nth0 < max_workgroup_size) {
+        nth0 *= 2;
+    }
+
+    int rows_per_workgroup = 1;
+    if (nth0 > nblk0) {
+        rows_per_workgroup = nth0 / nblk0;
+        nth0 = nblk0;
+    }
+
+    size_t global_work_size[] = {
+        (size_t)(ne01 + rows_per_workgroup - 1)/rows_per_workgroup*nth0,
+        (size_t)ne02*rows_per_workgroup,
+        (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth0, (size_t)rows_per_workgroup, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -4784,6 +4946,58 @@ static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
 }
 
+static void ggml_cl_mul_mat_f16_f32_tiled(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    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 * 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 offsetd = extrad->offset + dst->view_offs;
+
+    const int M = src0->ne[1];
+    const int N = src1->ne[1];
+    const int K = src0->ne[0];
+
+    cl_kernel kernel = backend_ctx->kernel_mul_mat_f16_f32_tiled;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(int),      &M));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(int),      &N));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),      &K));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &offsetd));
+
+    // Tiling parameters. These need to be tuned for optimal performance.
+    // They must match the #defines in the kernel mul_mat_f16_f32.cl.
+    //
+    // OPWM / OPWN: Output tile size per Work-Group. A work-group computes a tile of size OPWM x OPWN.
+    // TPWM / TPWN: Threads per Work-group. This is the work-group size.
+    // OPTM / OPTN: Output elements per Thread. Each thread computes OPTM x OPTN elements.
+    //
+    // The following relationships must hold:
+    //   OPWM = TPWM * OPTM
+    //   OPWN = TPWN * OPTN
+    //
+    const int OPWM = 64;
+    const int OPWN = 64;
+    const int TPWM = 16;
+    const int TPWN = 8;
+
+    size_t local_work_size[2] = { TPWM, TPWN };
+    size_t global_work_size[2] = {
+        (size_t) ((M + OPWM - 1) / OPWM) * TPWM,
+        (size_t) ((N + OPWN - 1) / OPWN) * TPWN,
+    };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -4797,6 +5011,18 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
+     if (src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32 &&
+        src0->ne[1] > 32 &&   // M > 32
+        src1->ne[1] > 32 &&   // N > 32
+        src0->ne[0] > 32 &&   // K > 32
+        src0->ne[2] == 1 && src0->ne[3] == 1 &&
+        src1->ne[2] == 1 && src1->ne[3] == 1 &&
+        ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
+        backend_ctx->kernel_mul_mat_f16_f32_tiled != NULL) {
+        ggml_cl_mul_mat_f16_f32_tiled(backend, src0, src1, dst);
+        return;
+    }
+
     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 * extrad = (ggml_tensor_extra_cl *)dst->extra;
@@ -6388,6 +6614,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_get_rows;
             break;
+        case GGML_OP_SET_ROWS:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_set_rows;
+            break;
         case GGML_OP_CPY:
             if (!any_on_device) {
                 return false;

ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl

@@ -0,0 +1,130 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define OPWM 64
+#define OPWN 64
+#define CPWK 8
+#define OPTM 4
+#define OPTN 8
+
+#define WG_M (OPWM / OPTM)
+#define WG_N (OPWN / OPTN)
+#define VEC_K (CPWK / 4)
+
+REQD_SUBGROUP_SIZE_128
+__kernel void mul_mat_f16_f32(
+    const int M, const int N, const int K,
+    __global const void* A_void, ulong A_offset,
+    __global const void* B_void, ulong B_offset,
+    __global       void* C_void, ulong C_offset) {
+
+    __global const half*  A = (__global const half* )((__global const char*)A_void + A_offset);
+    __global const float* B = (__global const float*)((__global const char*)B_void + B_offset);
+    __global       float* C = (__global       float*)((__global       char*)C_void + C_offset);
+
+    const int lidm = get_local_id(0);
+    const int lidn = get_local_id(1);
+    const int lid = lidn * WG_M + lidm;
+
+    const int offsetM = get_group_id(0) * OPWM;
+    const int offsetN = get_group_id(1) * OPWN;
+
+    __local half4  Alocal[OPWM][VEC_K];
+    __local float4 Blocal[OPWN][VEC_K];
+
+    float sum[OPTM][OPTN];
+
+    for (int wm = 0; wm < OPTM; wm++) {
+        for (int wn = 0; wn < OPTN; wn++) {
+            sum[wm][wn] = 0.0f;
+        }
+    }
+
+    const int numTiles = (K + CPWK - 1) / CPWK;
+
+    const int load_row_a = lid % OPWM;
+    const int load_vec_k_a = lid / OPWM;
+    const int global_row_a = offsetM + load_row_a;
+
+    const int load_row_b = lid % OPWN;
+    const int load_vec_k_b = lid / OPWN;
+    const int global_row_b = offsetN + load_row_b;
+
+    for (int t = 0; t < numTiles; t++) {
+        const int k_start = t * CPWK;
+        const int k_vec_start_a = k_start + load_vec_k_a * 4;
+        const int k_vec_start_b = k_start + load_vec_k_b * 4;
+
+        if (global_row_a < M && k_vec_start_a < K) {
+            if (k_vec_start_a + 3 < K) {
+                Alocal[load_row_a][load_vec_k_a] = vload4(0, A + global_row_a * K + k_vec_start_a);
+            } else {
+                half4 tempA = (half4)(0.0h);
+                if (k_vec_start_a < K) tempA.s0 = A[global_row_a * K + k_vec_start_a];
+                if (k_vec_start_a + 1 < K) tempA.s1 = A[global_row_a * K + k_vec_start_a + 1];
+                if (k_vec_start_a + 2 < K) tempA.s2 = A[global_row_a * K + k_vec_start_a + 2];
+                Alocal[load_row_a][load_vec_k_a] = tempA;
+            }
+        } else {
+            Alocal[load_row_a][load_vec_k_a] = (half4)(0.0h);
+        }
+
+        if (global_row_b < N && k_vec_start_b < K) {
+            if (k_vec_start_b + 3 < K) {
+                Blocal[load_row_b][load_vec_k_b] = vload4(0, B + global_row_b * K + k_vec_start_b);
+            } else {
+                float4 tempB = (float4)(0.0f);
+                if (k_vec_start_b < K) tempB.s0 = B[global_row_b * K + k_vec_start_b];
+                if (k_vec_start_b + 1 < K) tempB.s1 = B[global_row_b * K + k_vec_start_b + 1];
+                if (k_vec_start_b + 2 < K) tempB.s2 = B[global_row_b * K + k_vec_start_b + 2];
+                Blocal[load_row_b][load_vec_k_b] = tempB;
+            }
+        } else {
+            Blocal[load_row_b][load_vec_k_b] = (float4)(0.0f);
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (int k_vec = 0; k_vec < VEC_K; k_vec++) {
+            float4 a_fvecs[OPTM];
+            int current_row_a = lidm;
+            for (int wm = 0; wm < OPTM; wm++) {
+                a_fvecs[wm] = convert_float4(Alocal[current_row_a][k_vec]);
+                current_row_a += WG_M;
+            }
+
+            float4 b_fvecs[OPTN];
+            int current_row_b = lidn;
+            for (int wn = 0; wn < OPTN; wn++) {
+                b_fvecs[wn] = Blocal[current_row_b][k_vec];
+                current_row_b += WG_N;
+            }
+
+            for (int wm = 0; wm < OPTM; wm++) {
+                for (int wn = 0; wn < OPTN; wn++) {
+                    sum[wm][wn] += dot(a_fvecs[wm], b_fvecs[wn]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (int wm = 0; wm < OPTM; wm++) {
+        int globalRow = offsetM + lidm + wm * WG_M;
+        if (globalRow < M) {
+            for (int wn = 0; wn < OPTN; wn++) {
+                int globalCol = offsetN + lidn + wn * WG_N;
+                if (globalCol < N) {
+                    C[globalCol * M + globalRow] = sum[wm][wn];
+                }
+            }
+        }
+    }
+}

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

@@ -0,0 +1,95 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_set_rows_f32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        int           ne11,
+        int           ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = i03%ne12;
+    int i11 = i02%ne11;
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global float * dst_row = (global float *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = (float)src_row[ind];
+    }
+}
+
+kernel void kernel_set_rows_f16(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        int           ne11,
+        int           ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = i03%ne12;
+    int i11 = i02%ne11;
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global half  * dst_row = (global half  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = src_row[ind];
+    }
+}

ggml/src/ggml-sycl/backend.hpp

@@ -30,6 +30,7 @@
 #include "outprod.hpp"
 #include "quants.hpp"
 #include "rope.hpp"
+#include "set_rows.hpp"
 #include "softmax.hpp"
 #include "tsembd.hpp"
 #include "wkv.hpp"

ggml/src/ggml-sycl/gemm.hpp

@@ -32,39 +32,28 @@ public:
         else static_assert(0);
     }
 
-    // matrix A has m rows, k columns
-    // matrix B has k rows, n columns
-    // nra - number of elements to skip when moving into next row in A
-    // nrb - number of elements to skip when moving into next row in B
-    // nca - number of elements to skip when moving into next column in A
-    // ncb - number of elements to skip when moving into next column in B
-    // stride_a - number of elements to skip when moving to next A matrix
-    // stride_b - number of elements to skip when moving to next B matrix
-    // batches_a - number of A matrices
-    // batches_b - number of B matrices
     static void gemm(ggml_backend_sycl_context & ctx, int m, int n, int k,
-        const void * a, dt at, dnnl_dim_t nra, dnnl_dim_t nca, dnnl_dim_t stride_a,
-        const void * b, dt bt, dnnl_dim_t nrb, dnnl_dim_t ncb, dnnl_dim_t stride_b,
+        const void * a, dt at, dnnl_dim_t stra0, dnnl_dim_t stra1, dnnl_dim_t stra2,
+        const void * b, dt bt, dnnl_dim_t strb0, dnnl_dim_t strb1, dnnl_dim_t strb2,
         void * c, dt ct, const queue_ptr & q, dnnl_dim_t batches_a, dnnl_dim_t batches_b) {
 
         auto stream = ctx.stream_dnnl(q);
         auto eng = ctx.engine_dnnl(q);
 
-        // { # strides, # rows, # columns }
-        dnnl::memory::dims a_dims = { batches_a, m, k };
-        dnnl::memory::dims b_dims = { batches_b, k, n };
-        dnnl::memory::dims c_dims = { std::max(batches_a, batches_b), m, n };
-
-        // { # elements to skip to next stride, # elements to skip to next row, # elements to skip to next column }
-        dnnl::memory::dims a_strides = { stride_a, nra, nca };
-        dnnl::memory::dims b_strides = { stride_b, nrb, ncb };
-
+        dnnl::memory::dims a_dims = {batches_a, m, k };
+        dnnl::memory::dims a_strides = {stra2, stra1, stra0};
         const auto a_in_md = dnnl::memory::desc(a_dims, at, a_strides);
-        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_strides);
-        const auto c_md    = dnnl::memory::desc(c_dims, ct, tag::abc);
 
+        dnnl::memory::dims b_dims = {batches_b, k, n };
+        dnnl::memory::dims b_strides = {strb2, strb0, strb1};
+        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_strides);
+
+        dnnl::memory::dims c_dims = { std::max(batches_a, batches_b), m, n};
+        dnnl::memory::dims c_strides = {m*n, 1,  m };
+        const auto c_md    = dnnl::memory::desc(c_dims, ct, c_strides);
         dnnl::primitive_attr primitive_attr;
         primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
+
 #ifdef GGML_SYCL_F16
         primitive_attr.set_fpmath_mode(dnnl::fpmath_mode::f16);
 #endif
@@ -76,24 +65,23 @@ public:
 
         auto scratchpad_md = matmul_pd.scratchpad_desc();
         auto scratchpad_mem = ctx.get_scratchpad_mem(scratchpad_md, eng, q);
+
         auto matmul_prim = dnnl::matmul(matmul_pd);
 
         std::unordered_map<int, dnnl::memory> matmul_args;
         matmul_args.insert({ DNNL_ARG_SRC, a_mem });
         matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem });
+
         matmul_args.insert({ DNNL_ARG_DST, c_mem });
         matmul_args.insert({ DNNL_ARG_SCRATCHPAD, scratchpad_mem });
 
         matmul_prim.execute(stream, matmul_args);
     }
 
-    // matrices A and B are column major, both having k rows
-    // matrix A has m column, matrix B has n columns
-    // output: column major matrix C = A transposed * B
     static void row_gemm(ggml_backend_sycl_context & ctx, int m, int n, int k,
         const void * a, dt at, const void * b, dt bt, void * c, dt ct, const queue_ptr & q) {
 
-        gemm(ctx, m, n, k, a, at, k, 1, k * m, b, bt, 1, k, n * k, c, ct, q, 1, 1);
+        gemm(ctx, m, n, k, a, at, 1, k, k * m, b, bt, 1, k, n * k, c, ct, q, 1, 1);
     }
 };
 

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

@@ -41,6 +41,7 @@
 #include "ggml-sycl/element_wise.hpp"
 #include "ggml-sycl/presets.hpp"
 #include "ggml-sycl/gemm.hpp"
+#include "ggml-sycl/set_rows.hpp"
 #include "ggml-sycl/sycl_hw.hpp"
 #include "ggml-sycl/getrows.hpp"
 #include "ggml.h"
@@ -1545,7 +1546,7 @@ static void mul_mat_p021_f16_f32(
 
 static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
     const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor,
+    const int row_stride_x, const int channel_stride_x,const int channel_stride_y, const int channel_x_divisor,
     const sycl::nd_item<3> &item_ct1) {
 
     const sycl::half *x = (const sycl::half *)vx;
@@ -1556,7 +1557,6 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
                         item_ct1.get_local_id(0);
     const int channel_x = channel / channel_x_divisor;
 
-    const int nrows_y   = ncols_x;
     const int nrows_dst = nrows_x;
     const int row_dst   = row_x;
 
@@ -1575,7 +1575,7 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
         const int row_y = col_x;
 
         const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
-        const int iy = channel*nrows_y + row_y;
+        const int iy = channel * channel_stride_y + row_y;
 
         const float xi =
             sycl::vec<sycl::half, 1>(x[ix])
@@ -1822,7 +1822,7 @@ static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y,
 static void ggml_mul_mat_vec_nc_f16_f32_sycl(
     const void *vx, const float *y, float *dst, const int ncols_x,
     const int nrows_x, const int row_stride_x, const int nchannels_x,
-    const int nchannels_y, const int channel_stride_x, queue_ptr stream) {
+    const int nchannels_y, const int channel_stride_x, const int channel_stride_y, queue_ptr stream) {
 
     const sycl::range<3> block_nums(nchannels_y, nrows_x, 1);
     const sycl::range<3> block_dims(1, 1, WARP_SIZE);
@@ -1834,7 +1834,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl(
             sycl::nd_range<3>(block_nums * block_dims, block_dims),
             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 mul_mat_vec_nc_f16_f32(vx, y, dst, ncols_x, nrows_x,
-                                       row_stride_x, channel_stride_x,
+                                       row_stride_x, channel_stride_x, channel_stride_y,
                                        nchannels_y / nchannels_x, item_ct1);
             });
     }
@@ -2123,8 +2123,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
 
 #if GGML_SYCL_DNNL
         if (!g_ggml_sycl_disable_dnn) {
-            DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ptr,
-                                      DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
+                DnnlGemmWrapper::row_gemm(ctx,row_diff, src1_ncols , ne10, src0_ptr,
+                                     DnnlGemmWrapper::to_dt<sycl::half>(), src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
                                       dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
         }
         else
@@ -2170,8 +2170,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
 
 #if GGML_SYCL_DNNL
         if (!g_ggml_sycl_disable_dnn) {
-            DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ddf1_i,
-                                      DnnlGemmWrapper::to_dt<float>(), src0_ddf_i, DnnlGemmWrapper::to_dt<float>(),
+            DnnlGemmWrapper::row_gemm(ctx, row_diff, src1_ncols, ne10, src0_ddf_i,
+                                      DnnlGemmWrapper::to_dt<float>(), src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
                                       dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
         }
         else
@@ -2775,6 +2775,7 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
     const int64_t nb02 = src0->nb[2];
 
     const int64_t ne12 = src1->ne[2];
+    const int64_t nb11 = src1->nb[1];
 
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
     queue_ptr main_stream = ctx.stream();
@@ -2785,8 +2786,9 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
 
     const int64_t row_stride_x = nb01 / sizeof(sycl::half);
     const int64_t channel_stride_x = nb02 / sizeof(sycl::half);
+    const int64_t channel_stride_y = nb11 / sizeof(float);
 
-    ggml_mul_mat_vec_nc_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
+    ggml_mul_mat_vec_nc_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x,channel_stride_y, main_stream);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -2840,8 +2842,8 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     float *            dst_ddf  = static_cast<float *>(dst->data);
 
     const sycl::half * src1_f16       = static_cast<const sycl::half *>(src1->data);
+    const size_t       type_size_src0 = ggml_type_size(src0->type);
     const size_t       type_size_src1 = ggml_type_size(src1->type);
-    GGML_ASSERT(nb10 == type_size_src1);
 
     // SRC1 strides
     int64_t                          s11 = nb11 / type_size_src1;
@@ -2853,11 +2855,40 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     if (src1->type != GGML_TYPE_F16) {
         scope_op_debug_print    scope_dbg_print(__func__, "/to_fp16_nc_sycl", dst, /*num_src=*/2,
                                                 " : converting src1 to fp16");
-        const to_fp16_nc_sycl_t to_fp16_nc_sycl = get_to_fp16_nc_sycl(src1->type);
-        GGML_ASSERT(to_fp16_nc_sycl != nullptr);
+
+        // iterate tensor dims and find the slowest moving dim and stride
+        int64_t last_dim=0;
+        int64_t last_str=0;
+        int64_t largest_str=0;
+        for(int i = 0; i< 4; i++){
+            // last stride is always the largest
+            if(src1->nb[i] == largest_str){
+                if(src1->ne[last_dim] == 1){
+                    last_str = i;
+                    last_dim = i;
+                }
+            }
+            if(src1->nb[i] > largest_str){
+                largest_str = src1->nb[i];
+                last_str = i;
+                last_dim = i;
+            }
+
+        }
+#if GGML_SYCL_DNNL
+        // oneDNN handles strided data and does not need overhead of get_to_fp16_nc_sycl
+        const int64_t ne_src1 = src1->nb[last_str] * src1->ne[last_dim] / type_size_src1;
+        src1_f16_alloc.alloc(ne_src1);
+        const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst);
+        GGML_ASSERT(to_fp16_sycl != nullptr);
+        to_fp16_sycl(src1_f16, src1_f16_alloc.get(), ne_src1, queue);
+# else
         const int64_t ne_src1 = ggml_nelements(src1);
         src1_f16_alloc.alloc(ne_src1);
+        const to_fp16_nc_sycl_t to_fp16_nc_sycl = get_to_fp16_nc_sycl(src1->type);
+        GGML_ASSERT(to_fp16_nc_sycl != nullptr);
         to_fp16_nc_sycl(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, queue);
+#endif
 
         src1_f16 = src1_f16_alloc.get();
         s11      = ne10;
@@ -2891,38 +2922,89 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
 
 #if GGML_SYCL_DNNL
     if (!g_ggml_sycl_disable_dnn) {
-        auto dnn_gemm = [&ctx, queue, ne11, ne01, ne10, nb00, nb01, nb02, s11, s12]
-            (const sycl::half* src1, const sycl::half* src0, float* dst, const dnnl_dim_t batches_a, const dnnl_dim_t batches_b) {
+            int64_t str_a0 = nb00 / type_size_src0;
+            int64_t str_a1 = nb01 / type_size_src0;
+            int64_t str_a2 = nb02 / type_size_src0;
 
-            DnnlGemmWrapper::gemm(ctx, ne11,ne01, ne10,
-                            src1, DnnlGemmWrapper::to_dt<sycl::half>(), s11, 1, s12,
-                            src0, DnnlGemmWrapper::to_dt<sycl::half>(), 1, nb01/nb00, nb02/nb00,
-                            dst, DnnlGemmWrapper::to_dt<float>(), queue, batches_a, batches_b);
-        };
+            int64_t str_b0 = nb10 / type_size_src1;
+            int64_t str_b1 = nb11 / type_size_src1;
+            int64_t str_b2 = nb12 / type_size_src1;
 
-        if (r2 == 1 && r3 == 1) {
-            if (ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
-                dnn_gemm(src1_f16, src0_f16, dst_ddf, ne12*ne13, ne02 * ne03);
-            }
-            else {
-                for (int64_t ie03 = 0; ie03 < ne03; ++ie03) {
-                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie03*nb03)/sizeof(sycl::half)); // nb is in bytes
-                    const sycl::half* src1_f16_shifted = src1_f16 + ie03*s13;
-                    float* dst_shifted = dst_ddf + ((ie03*nb3)/sizeof(float));
-                    dnn_gemm(src1_f16_shifted, src0_f16_shifted, dst_shifted, ne12, ne02);
+            auto launch_gemm_for_batches = [&ctx, queue](const sycl::half *src0,
+                                                const sycl::half *src1, float *dst,
+                                                int64_t a0, int64_t a1, int64_t batcha,
+                                                int64_t b0, int64_t b1, int64_t batchb,
+                                                int64_t sa0, int64_t sa1, int64_t sa2,
+                                                int64_t sb0, int64_t sb1, int64_t sb2,
+                                                int64_t sd2) {
+                bool supported_broadcast = batchb == batcha ? true
+                        : batchb == 1 || batcha == 1        ? true
+                                                            : false;
+                if (supported_broadcast) {
+                    DnnlGemmWrapper::gemm(ctx, a1, b1, a0, src0,
+                            DnnlGemmWrapper::to_dt<sycl::half>(), sa0, sa1, sa2, src1,
+                            DnnlGemmWrapper::to_dt<sycl::half>(), sb0, sb1, sb2, dst,
+                            DnnlGemmWrapper::to_dt<float>(), queue, batcha, batchb);
+                } else {
+                    // iterate over batches from smaller set of matrices (matrix 0)
+                    int64_t batches0 = batcha;
+                    int64_t batches1 = batchb;
+
+                    if (batches0 > batches1) {
+                        int64_t num_mul_mats = batches1;
+                        int64_t sub_batch = batches0 / num_mul_mats;
+                        // src0 is batched and bigger, shift and multiply with src1
+                        for (int64_t i0 = 0; i0 < num_mul_mats; i0++) {
+                            const sycl::half *src0_shifted = src0 + (sa2 * i0 * sub_batch);
+                            const sycl::half *src1_shifted = src1 + (sb2 * i0);
+                            float *dst_shifted = dst + (sd2 * i0 * sub_batch);
+                            DnnlGemmWrapper::gemm(ctx, a1, b1, a0, src0_shifted,
+                                    DnnlGemmWrapper::to_dt<sycl::half>(), sa0, sa1, sa2,
+                                    src1_shifted, DnnlGemmWrapper::to_dt<sycl::half>(), sb0,
+                                    sb1, sb2, dst_shifted, DnnlGemmWrapper::to_dt<float>(),
+                                    queue, sub_batch, 1);
+                        }
+                    } else {
+                        int64_t num_mul_mats = batches0;
+                        int64_t sub_batch = batches1 / num_mul_mats;
+                        // src1 is batched and bigger, shift and multiply with src0
+                        for (int64_t i1 = 0; i1 < num_mul_mats; i1++) {
+                            const sycl::half *src0_shifted = src0 + (sa2 * i1);
+                            const sycl::half *src1_shifted = src1 + (sb2 * i1 * sub_batch);
+                            float *dst_shifted = dst + (sd2 * i1 * sub_batch);
+                            DnnlGemmWrapper::gemm(ctx, a1, b1, a0, src0_shifted,
+                                    DnnlGemmWrapper::to_dt<sycl::half>(), sa0, sa1, sa2,
+                                    src1_shifted, DnnlGemmWrapper::to_dt<sycl::half>(), sb0,
+                                    sb1, sb2, dst_shifted, DnnlGemmWrapper::to_dt<float>(),
+                                    queue, 1, sub_batch);
+                        }
+                    }
+                }
+            };
+
+            bool cont_batches_a = nb02 * ne02 == nb03;
+            bool cont_batches_b = nb12 * ne12 == nb13;
+            if (cont_batches_a && cont_batches_b) {
+                int64_t batches0 = ne02 * ne03;
+                int64_t batches1 = ne12 * ne13;
+                launch_gemm_for_batches(src0_f16, src1_f16, dst_ddf, ne00, ne01, batches0,
+                        ne10, ne11, batches1, str_a0, str_a1, str_a2, str_b0, str_b1,
+                        str_b2, nb2 / sizeof(float));
+            } else {
+                for (int64_t b_a = 0; b_a < ne03; b_a++) {
+                    const sycl::half *src0_f16_shifted
+                            = src0_f16 + (nb03 * b_a / type_size_src0);
+                    const sycl::half *src1_f16_shifted
+                            = src1_f16 + (nb13 * b_a / type_size_src1);
+                    float *dst_shifted = dst_ddf + (nb3 * b_a / sizeof(float));
+                    int64_t batches0 = ne02;
+                    int64_t batches1 = ne12;
+                    launch_gemm_for_batches(src0_f16_shifted, src1_f16_shifted, dst_shifted,
+                            ne00, ne01, batches0, ne10, ne11, batches1, str_a0, str_a1,
+                            str_a2, str_b0, str_b1, str_b2, nb2 / sizeof(float));
                 }
             }
-        } else {
-            // iterate over batches from smaller set of matrices (matrix 0)
-            for (int64_t ie02 = 0; ie02 < ne02; ++ie02) {
-                for (int64_t ie03 = 0; ie03 < ne03; ++ie03) {
-                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie02*nb02 + ie03*nb03)/sizeof(sycl::half));
-                    const sycl::half* src1_f16_shifted = src1_f16 + ie02*s12*r2 + ie03*s13*r3;
-                    float* dst_shifted = dst_ddf + ((ie02*nb2*r2 + ie03*nb3*r3)/sizeof(float));
-                    dnn_gemm(src1_f16_shifted, src0_f16_shifted, dst_shifted, r2*r3, 1);
-                }
-            }
-        }
+
     }
     else
 #endif
@@ -3262,10 +3344,10 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
             // The kernel from the if path is faster for that specific case, but does not support all mul mats.
             ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
         }
-    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2] * src1->ne[3] > 1) {
         // KQ + KQV multi-batch
         ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
@@ -3605,6 +3687,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_GET_ROWS:
             ggml_sycl_get_rows(ctx, dst);
             break;
+        case GGML_OP_SET_ROWS:
+            ggml_sycl_op_set_rows(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_sycl_dup(ctx, dst);
             break;
@@ -4299,7 +4384,8 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 // TODO: add support
                 // ref: https://github.com/ggml-org/llama.cpp/pull/14274
-                return false;
+#pragma message("TODO: implement BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
+                return (op->type == GGML_TYPE_F32 || (op->type == GGML_TYPE_F16 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_I64));
             } break;
         case GGML_OP_CPY:
             {

ggml/src/ggml-sycl/set_rows.cpp

@@ -0,0 +1,131 @@
+#include "set_rows.hpp"
+
+namespace utils {
+template<typename T>
+static constexpr bool is_arithmetic_v() {
+    return std::is_arithmetic_v<T> || std::is_same_v<T, sycl::half> || std::is_same_v<T, sycl::ext::oneapi::bfloat16>;
+}
+}
+
+template<typename TIn, typename TOut>
+static inline std::enable_if_t<utils::is_arithmetic_v<TIn>() && utils::is_arithmetic_v<TOut>(), void>
+convert (const char* src, char* dst) {
+    auto src_val = *reinterpret_cast<const TIn*>(src);
+    auto dst_val = sycl::vec<TIn, 1>(src_val).template convert<TOut, sycl::rounding_mode::automatic>()[0];
+   *reinterpret_cast<TOut*>(dst) = dst_val;
+}
+
+template<typename TIn, typename TOut>
+static void k_set_rows(
+        const char * __restrict__ src0, const int64_t * __restrict__ src1, char * __restrict__ dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t ne11, const int64_t ne12,
+        const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        const size_t src_type_size, const size_t dst_type_size,
+        const int64_t total_elements,
+        const sycl::nd_item<1> & item_ct1) {
+
+    const int64_t i = item_ct1.get_global_linear_id();
+    if (i >= total_elements) {
+        return;
+    }
+
+    const int64_t i03 = i / (ne00 * ne01 * ne02);
+    const int64_t i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int64_t i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
+    const int64_t i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+
+    const int64_t i12 = i03 % ne12;
+    const int64_t i11 = i02 % ne11;
+    const int64_t i10 = i01;
+
+    const int64_t dst_row = *(const int64_t *)((const char *)src1 + calculate_offset<3>({nb10, nb11, nb12}, {i10, i11, i12}));
+
+    const char * src0_row = src0 + calculate_offset<3>({nb01, nb02, nb03}, {i01, i02, i03});
+    const char * src_elem = src0_row + i00 * src_type_size;
+    char * dst_row_ptr = dst + dst_row*nb1 + i02*nb2 + i03*nb3;
+    char * dst_elem = dst_row_ptr + i00 * dst_type_size;
+
+    convert<TIn, TOut>(src_elem, dst_elem);
+}
+
+template<typename TIn, typename TOut>
+static void set_rows_sycl(
+        const char * src0_d, const int64_t * src1_d, char * dst_d,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne11, const int64_t ne12, const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        const size_t src_type_size, const size_t dst_type_size,
+        queue_ptr stream) {
+
+    const int64_t total_elements = ne00 * ne01 * ne02 * ne03;
+
+    constexpr int block_size = 64;
+    const int64_t grid_size = ceil_div(total_elements, block_size);
+
+    sycl_parallel_for(
+        stream,
+        sycl::nd_range<1>(grid_size * block_size, block_size),
+        [=](sycl::nd_item<1> item_ct1) {
+            k_set_rows<TIn, TOut>(
+                src0_d, src1_d, dst_d,
+                ne00, ne01, ne02,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                src_type_size, dst_type_size,
+                total_elements,
+                item_ct1
+            );
+        }
+    );
+}
+
+void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->src[1]->type == GGML_TYPE_I64);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int64_t * src1_dd = static_cast<const int64_t *>(src1->data);
+
+    dpct::queue_ptr stream = ctx.stream();
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            set_rows_sycl<float, float>(
+                (const char *)src0->data, src1_dd, (char *)dst->data,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                sizeof(float), sizeof(float),
+                stream
+            );
+            break;
+        case GGML_TYPE_F16:
+            dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+            set_rows_sycl<float, sycl::half>(
+                (const char *)src0->data, src1_dd, (char *)dst->data,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                sizeof(float), sizeof(sycl::half),
+                stream
+            );
+            break;
+        default:
+            GGML_ABORT("Unsupported tensor type!");
+            break;
+    }
+}

ggml/src/ggml-sycl/set_rows.hpp

@@ -0,0 +1,8 @@
+#ifndef GGML_SYCL_SET_ROWS_HPP
+#define GGML_SYCL_SET_ROWS_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_SET_ROWS_HPP

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

@@ -425,18 +425,20 @@ struct vk_device_struct {
     vk_pipeline pipeline_div_norepeat[2][2][2];
 
     vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
-    vk_pipeline pipeline_upscale_f32;
+    vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bilinear_ac_f32;
     vk_pipeline pipeline_scale_f32;
     vk_pipeline pipeline_sqr_f32;
     vk_pipeline pipeline_sin_f32;
     vk_pipeline pipeline_cos_f32;
     vk_pipeline pipeline_clamp_f32;
     vk_pipeline pipeline_pad_f32;
+    vk_pipeline pipeline_roll_f32;
     vk_pipeline pipeline_repeat_f32, pipeline_repeat_back_f32;
     vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16, pipeline_cpy_f16_f32, pipeline_cpy_f32_bf16;
     vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16, pipeline_contig_cpy_f16_f32, pipeline_contig_cpy_f32_bf16;
     vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
     vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_set_rows[GGML_TYPE_COUNT];
     vk_pipeline pipeline_norm_f32;
     vk_pipeline pipeline_group_norm_f32;
     vk_pipeline pipeline_rms_norm_f32;
@@ -693,6 +695,37 @@ struct vk_op_unary_push_constants {
 };
 static_assert(sizeof(vk_op_unary_push_constants) <= 128, "sizeof(vk_op_unary_push_constants) must be <= 128");
 
+static vk_op_unary_push_constants vk_op_unary_push_constants_init(const ggml_tensor * src0, const ggml_tensor * dst, int64_t ne = 0) {
+    GGML_ASSERT(ne != 0 || (ggml_nelements(src0) == ggml_nelements(dst)));
+    ne = ne != 0 ? ne : ggml_nelements(dst);
+    GGML_ASSERT(ne <= (int64_t)std::numeric_limits<uint32_t>::max());
+
+    vk_op_unary_push_constants p{};
+    p.ne = (uint32_t)ne;
+
+    size_t src0_tsize = ggml_type_size(src0->type);
+    p.ne00 = (uint32_t)src0->ne[0];
+    p.ne01 = (uint32_t)src0->ne[1];
+    p.ne02 = (uint32_t)src0->ne[2];
+    p.ne03 = (uint32_t)src0->ne[3];
+    p.nb00 = (uint32_t)(src0->nb[0] / src0_tsize);
+    p.nb01 = (uint32_t)(src0->nb[1] / src0_tsize);
+    p.nb02 = (uint32_t)(src0->nb[2] / src0_tsize);
+    p.nb03 = (uint32_t)(src0->nb[3] / src0_tsize);
+
+    size_t dst_tsize = ggml_type_size(dst->type);
+    p.ne10 = (uint32_t)dst->ne[0];
+    p.ne11 = (uint32_t)dst->ne[1];
+    p.ne12 = (uint32_t)dst->ne[2];
+    p.ne13 = (uint32_t)dst->ne[3];
+    p.nb10 = (uint32_t)(dst->nb[0] / dst_tsize);
+    p.nb11 = (uint32_t)(dst->nb[1] / dst_tsize);
+    p.nb12 = (uint32_t)(dst->nb[2] / dst_tsize);
+    p.nb13 = (uint32_t)(dst->nb[3] / dst_tsize);
+
+    return p; // fastdiv values and offsets are initialized later in ggml_vk_op
+}
+
 // See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
 // Precompute mp (m' in the paper) and L such that division
 // can be computed using a multiply (high 32b of 64b result)
@@ -862,6 +895,7 @@ struct vk_op_conv2d_dw_push_constants {
 
 struct vk_op_upscale_push_constants {
     uint32_t ne; uint32_t a_offset; uint32_t d_offset;
+    uint32_t ne00; uint32_t ne01;
     uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13;
     float sf0; float sf1; float sf2; float sf3;
@@ -1735,7 +1769,14 @@ static FaHeadSizes fa_get_head_sizes(uint32_t hsk, uint32_t hsv) {
 // number of rows/cols for flash attention shader
 static constexpr uint32_t flash_attention_num_small_rows = 32;
 static constexpr uint32_t scalar_flash_attention_num_small_rows = 1;
-static constexpr uint32_t scalar_flash_attention_num_large_rows = 8;
+
+static uint32_t get_fa_scalar_num_large_rows(uint32_t hsv) {
+    if (hsv >= 512) {
+        return 2;
+    } else {
+        return 8;
+    }
+}
 
 // The FA coopmat1 shader assumes 16x16x16 matrix multiply support.
 // 128 threads split into four subgroups, each subgroup does 1/4
@@ -1760,7 +1801,7 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint3
         if (small_rows) {
             return {scalar_flash_attention_num_small_rows, 64};
         } else {
-            return {scalar_flash_attention_num_large_rows, 32};
+            return {get_fa_scalar_num_large_rows(hsv), 32};
         }
     }
 
@@ -1779,7 +1820,11 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint3
 
     // small cols to reduce register count
     if (ggml_is_quantized(type) || hsk >= 256) {
-        return {64, 32};
+        if (hsk >= 512) {
+            return {32, 32};
+        } else {
+            return {64, 32};
+        }
     }
     return {64, 64};
 }
@@ -1821,7 +1866,7 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
     const uint32_t warps = warptile[0] / warptile[10];
 
     const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size;
-    const uint32_t mmid_row_ids = mul_mat_id ? 4096 * sizeof(uint32_t) : 0;
+    const uint32_t mmid_row_ids = mul_mat_id ? (4096 * sizeof(uint32_t) + 4/*_ne1*/) : 0;
     const uint32_t coopmat_stage = device->coopmat_support ? warptile[7] * warptile[8] / warps * sizeof(float) : 0;
 
     const uint32_t total_size = load_bufs + mmid_row_ids + coopmat_stage + lut_size;
@@ -1946,10 +1991,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
         s_mmq_wg_denoms_k = { 32,  32, 1 };
 
         // spec constants and tile sizes for quant matmul_id
-        l_warptile_mmqid = { 256, 128, 64, 16, 0 };
+        l_warptile_mmqid = { 256, 128, 128, 16, 0 };
         m_warptile_mmqid = { 256, 128, 64, 16, 0 };
         s_warptile_mmqid = { 256, 128, 64, 16, 0 };
-        l_mmqid_wg_denoms = { 128, 64, 1 };
+        l_mmqid_wg_denoms = { 128, 128, 1 };
         m_mmqid_wg_denoms = { 128, 64, 1 };
         s_mmqid_wg_denoms = { 128, 64, 1 };
 
@@ -2738,19 +2783,41 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_bf16,"contig_cpy_f32_bf16",contig_cpy_f32_bf16_len,contig_cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     if (device->float_controls_rte_fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_rte_len, cpy_f32_q5_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_rte_len, cpy_f32_q5_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_rte_len, cpy_f32_q8_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_rte_len, cpy_f32_iq4_nl_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_rte_len, cpy_f32_q5_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_rte_len, cpy_f32_q5_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_rte_len, cpy_f32_q8_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_rte_len, cpy_f32_iq4_nl_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
     } else {
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
+    }
+
+    if (device->float_controls_rte_fp16) {
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F32],  "set_rows_f32",  set_rows_f32_rte_len,  set_rows_f32_rte_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F16],  "set_rows_f16",  set_rows_f16_rte_len,  set_rows_f16_rte_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_BF16], "set_rows_bf16", set_rows_bf16_rte_len, set_rows_bf16_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_0], "set_rows_q4_0", set_rows_q4_0_rte_len, set_rows_q4_0_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_1], "set_rows_q4_1", set_rows_q4_1_rte_len, set_rows_q4_1_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_0], "set_rows_q5_0", set_rows_q5_0_rte_len, set_rows_q5_0_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_1], "set_rows_q5_1", set_rows_q5_1_rte_len, set_rows_q5_1_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q8_0], "set_rows_q8_0", set_rows_q8_0_rte_len, set_rows_q8_0_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_IQ4_NL], "set_rows_iq4_nl", set_rows_iq4_nl_rte_len, set_rows_iq4_nl_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+    } else {
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F32],  "set_rows_f32",  set_rows_f32_len,  set_rows_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F16],  "set_rows_f16",  set_rows_f16_len,  set_rows_f16_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_BF16], "set_rows_bf16", set_rows_bf16_len, set_rows_bf16_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_0], "set_rows_q4_0", set_rows_q4_0_len, set_rows_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_1], "set_rows_q4_1", set_rows_q4_1_len, set_rows_q4_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_0], "set_rows_q5_0", set_rows_q5_0_len, set_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_1], "set_rows_q5_1", set_rows_q5_1_len, set_rows_q5_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q8_0], "set_rows_q8_0", set_rows_q8_0_len, set_rows_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_IQ4_NL], "set_rows_iq4_nl", set_rows_iq4_nl_len, set_rows_iq4_nl_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_0], "cpy_q4_0_f32", cpy_q4_0_f32_len, cpy_q4_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
@@ -2790,7 +2857,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_concat_i32, "concat_i32", concat_i32_len, concat_i32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_upscale_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_upscale_nearest_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_NEAREST}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_ac_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ALIGN_CORNERS}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -2802,6 +2871,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_roll_f32, "roll_f32", roll_f32_len, roll_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(device, device->pipeline_repeat_f32, "repeat_f32", repeat_f32_len, repeat_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_repeat_back_f32, "repeat_back_f32", repeat_back_f32_len, repeat_back_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -6048,7 +6119,7 @@ static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, con
     // Needs to be kept up to date on shader changes
     GGML_UNUSED(hsv);
     const uint32_t wg_size = scalar_flash_attention_workgroup_size;
-    const uint32_t Br = scalar_flash_attention_num_large_rows;
+    const uint32_t Br = get_fa_scalar_num_large_rows(hsv);
     const uint32_t Bc = scalar_flash_attention_Bc;
 
     const uint32_t tmpsh = wg_size * sizeof(float);
@@ -6173,7 +6244,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     case FA_SCALAR:
     case FA_COOPMAT1:
         // We may switch from coopmat1 to scalar, so use the scalar limit for both
-        max_gqa = scalar_flash_attention_num_large_rows;
+        max_gqa = get_fa_scalar_num_large_rows(HSV);
         break;
     case FA_COOPMAT2:
         max_gqa = get_fa_num_small_rows(FA_COOPMAT2);
@@ -6468,8 +6539,16 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_UPSCALE:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && dst->op_params[0] == GGML_SCALE_MODE_NEAREST) {
-            return ctx->device->pipeline_upscale_f32;
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            int mode = ggml_get_op_params_i32(dst, 0);
+            switch (mode) {
+                case GGML_SCALE_MODE_NEAREST:
+                    return ctx->device->pipeline_upscale_nearest_f32;
+                case GGML_SCALE_MODE_BILINEAR:
+                    return ctx->device->pipeline_upscale_bilinear_f32;
+                case GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ALIGN_CORNERS:
+                    return ctx->device->pipeline_upscale_bilinear_ac_f32;
+            }
         }
         return nullptr;
     case GGML_OP_SCALE:
@@ -6502,6 +6581,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_pad_f32;
         }
         return nullptr;
+    case GGML_OP_ROLL:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_roll_f32;
+        }
+        return nullptr;
     case GGML_OP_REPEAT:
         if (ggml_type_size(src0->type) == sizeof(float) && ggml_type_size(dst->type) == sizeof(float)) {
             return ctx->device->pipeline_repeat_f32;
@@ -6516,6 +6600,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     case GGML_OP_CONT:
     case GGML_OP_DUP:
         return ggml_vk_get_cpy_pipeline(ctx, src0, dst, dst->type);
+    case GGML_OP_SET_ROWS:
+        return ctx->device->pipeline_set_rows[dst->type];
     case GGML_OP_SILU_BACK:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_silu_back_f32;
@@ -6754,6 +6840,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_RMS_NORM:
     case GGML_OP_CONV_2D_DW:
     case GGML_OP_IM2COL:
+    case GGML_OP_SET_ROWS:
         return true;
     default:
         return false;
@@ -7048,6 +7135,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
+    case GGML_OP_ROLL:
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_CPY:
@@ -7067,6 +7155,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
                     ne *= ggml_type_size(src0->type) / 2;
                 }
             }
+            // copy_to_quant has block size of 32, and each thread does QUANT_K elements.
+            // Splitting into 512x512xZ wouldn't work well since each workgroup does 1024 elements.
+            // So divide by block size here before splitting into 512x512 groups.
+            if (op == GGML_OP_CPY && !ggml_is_quantized(src0->type) && ggml_is_quantized(dst->type)) {
+                ne = CEIL_DIV(ne, ggml_blck_size(dst->type));
+            }
             if (ne > 262144) {
                 elements = { 512, 512, CEIL_DIV(ne, 262144) };
             } else if (ne > 512) {
@@ -7075,6 +7169,25 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
                 elements = { ne, 1, 1 };
             }
         } break;
+    case GGML_OP_SET_ROWS:
+        {
+            uint32_t ne = ggml_nelements(src0);
+            if (ggml_is_quantized(dst->type)) {
+                // quants run 32 threads each doing QUANT_K elements
+                ne = CEIL_DIV(ne, 32 * ggml_blck_size(dst->type));
+            } else {
+                // scalar types do one element per thread, running 512 threads
+                ne = CEIL_DIV(ne, 512);
+            }
+            if (ne > 262144) {
+                elements = { 512, 512, CEIL_DIV(ne, 262144) };
+            } else if (ne > 512) {
+                elements = { 512, CEIL_DIV(ne, 512), 1 };
+            } else {
+                elements = { ne, 1, 1 };
+            }
+        }
+        break;
     default:
         elements = { (uint32_t)ggml_nelements(src0), 1, 1 };
         break;
@@ -7484,14 +7597,21 @@ static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, co
 
 static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t mode = (uint32_t)ggml_get_op_params_i32(dst, 0);
 
-    const float sf0 = (float)dst->ne[0] / src0->ne[0];
-    const float sf1 = (float)dst->ne[1] / src0->ne[1];
-    const float sf2 = (float)dst->ne[2] / src0->ne[2];
-    const float sf3 = (float)dst->ne[3] / src0->ne[3];
+    float sf0 = (float)dst->ne[0] / src0->ne[0];
+    float sf1 = (float)dst->ne[1] / src0->ne[1];
+    float sf2 = (float)dst->ne[2] / src0->ne[2];
+    float sf3 = (float)dst->ne[3] / src0->ne[3];
+
+    if (mode & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+        sf0 = (float)(dst->ne[0] - 1) / (src0->ne[0] - 1);
+        sf1 = (float)(dst->ne[1] - 1) / (src0->ne[1] - 1);
+    }
 
     ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
         (uint32_t)ggml_nelements(dst), 0, 0,
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1],
         (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], (uint32_t)dst->ne[2],(uint32_t)dst->ne[3],
         sf0, sf1, sf2, sf3,
@@ -7499,123 +7619,64 @@ static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, c
 }
 
 static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    float * op_params = (float *)dst->op_params;
-    const uint32_t src0_type_size = ggml_type_size(src0->type);
-    const uint32_t dst_type_size = ggml_type_size(dst->type);
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = ggml_get_op_params_f32(dst, 0);
+    p.param2 = ggml_get_op_params_f32(dst, 1);
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SCALE, {
-        (uint32_t)ggml_nelements(src0),
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        0,
-        op_params[0], op_params[1],
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    }, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SCALE, std::move(p), dryrun);
 }
 
 static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    const uint32_t src0_type_size = ggml_type_size(src0->type);
-    const uint32_t dst_type_size = ggml_type_size(dst->type);
-
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, {
-        (uint32_t)ggml_nelements(src0),
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        0,
-        0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    }, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
 
 static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    const uint32_t src0_type_size = ggml_type_size(src0->type);
-    const uint32_t dst_type_size = ggml_type_size(dst->type);
-
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SIN, {
-        (uint32_t)ggml_nelements(src0),
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        0,
-        0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    }, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
 
 static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    const uint32_t src0_type_size = ggml_type_size(src0->type);
-    const uint32_t dst_type_size = ggml_type_size(dst->type);
-
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_COS, {
-        (uint32_t)ggml_nelements(src0),
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        0,
-        0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    }, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_COS, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
 
 static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    float * op_params = (float *)dst->op_params;
-    const uint32_t src0_type_size = ggml_type_size(src0->type);
-    const uint32_t dst_type_size = ggml_type_size(dst->type);
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = ggml_get_op_params_f32(dst, 0);
+    p.param2 = ggml_get_op_params_f32(dst, 1);
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CLAMP, {
-        (uint32_t)ggml_nelements(src0),
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        0,
-        op_params[0], op_params[1],
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    }, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CLAMP, std::move(p), dryrun);
 }
 
 static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    const uint32_t src0_type_size = ggml_type_size(src0->type);
-    const uint32_t dst_type_size = ggml_type_size(dst->type);
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p), dryrun);
+}
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_PAD, {
-        (uint32_t)ggml_nelements(dst),
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        0,
-        0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    }, dryrun);
+static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const int32_t s0 = ggml_get_op_params_i32(dst, 0);
+    const int32_t s1 = ggml_get_op_params_i32(dst, 1);
+    const int32_t s2 = ggml_get_op_params_i32(dst, 2);
+    const int32_t s3 = ggml_get_op_params_i32(dst, 3);
+    const uint32_t s01_packed = ((s0 + 0x8000) << 16) | (s1 + 0x8000);
+    const uint32_t s23_packed = ((s2 + 0x8000) << 16) | (s3 + 0x8000);
+
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    memcpy(&p.param1, &s01_packed, sizeof(float));
+    memcpy(&p.param2, &s23_packed, sizeof(float));
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ROLL, std::move(p), dryrun);
 }
 
 static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    const uint32_t src0_type_size = ggml_type_size(src0->type);
-    const uint32_t dst_type_size = ggml_type_size(dst->type);
-
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT, {
-        (uint32_t)ggml_nelements(dst),
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        0,
-        0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    }, dryrun);
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT, std::move(p), dryrun);
 }
 
 static void ggml_vk_repeat_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    const uint32_t src0_type_size = ggml_type_size(src0->type);
-    const uint32_t dst_type_size = ggml_type_size(dst->type);
-
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, {
-        (uint32_t)ggml_nelements(dst),
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        0,
-        0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    }, dryrun);
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, std::move(p), dryrun);
 }
 
 static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    const uint32_t src0_type_size = ggml_type_size(src0->type);
-    const uint32_t dst_type_size = ggml_type_size(dst->type);
-
     uint32_t ne = (uint32_t)ggml_nelements(src0);
     if (ggml_is_quantized(src0->type) && ggml_is_quantized(dst->type)) {
         // Convert from number of logical elements to 2- or 4-byte units.
@@ -7627,13 +7688,22 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const
         }
     }
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, {
-        ne,
-        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
-        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ne);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, std::move(p), dryrun);
+}
+
+static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SET_ROWS, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        0.0f, 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+        0.0f, 0.0f, 0,
     }, dryrun);
 }
 
@@ -8956,7 +9026,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
+    case GGML_OP_ROLL:
     case GGML_OP_CPY:
+    case GGML_OP_SET_ROWS:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
     case GGML_OP_SILU_BACK:
@@ -9023,6 +9095,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_CLAMP:
         case GGML_OP_PAD:
         case GGML_OP_CPY:
+        case GGML_OP_SET_ROWS:
         case GGML_OP_CONT:
         case GGML_OP_DUP:
         case GGML_OP_SILU_BACK:
@@ -9125,12 +9198,20 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_PAD:
         ggml_vk_pad(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_ROLL:
+        ggml_vk_roll(ctx, compute_ctx, src0, node, dryrun);
+
         break;
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
         ggml_vk_cpy(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_SET_ROWS:
+        ggml_vk_set_rows(ctx, compute_ctx, src0, src1, node, dryrun);
+
         break;
     case GGML_OP_SILU_BACK:
         ggml_vk_silu_back(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -9345,7 +9426,9 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
+    case GGML_OP_ROLL:
     case GGML_OP_CPY:
+    case GGML_OP_SET_ROWS:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
     case GGML_OP_SILU_BACK:
@@ -10411,9 +10494,20 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             } break;
         case GGML_OP_SET_ROWS:
             {
-                // TODO: add support
-                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
-                return false;
+                switch (op->type) {
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
+                    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_IQ4_NL:
+                        return true;
+                    default:
+                        return false;
+                }
             } break;
         case GGML_OP_CONT:
         case GGML_OP_CPY:
@@ -10499,13 +10593,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_CLAMP:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_UPSCALE:
-            return op->op_params[0] == GGML_SCALE_MODE_NEAREST;
         case GGML_OP_ACC:
         case GGML_OP_CONCAT:
         case GGML_OP_SCALE:
         case GGML_OP_PAD:
+        case GGML_OP_ROLL:
         case GGML_OP_DIAG_MASK_INF:
-            return true;
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
         case GGML_OP_ARGSORT:
@@ -11028,6 +11121,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else {
             tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]);
         }
+    } else if (tensor->op == GGML_OP_SET_ROWS) {
+        tensor_clone = ggml_set_rows(ggml_ctx, src_clone[0], src_clone[1]);
     } else if (tensor->op == GGML_OP_CONT) {
         tensor_clone = ggml_cont_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
     } else if (tensor->op == GGML_OP_RESHAPE) {

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

@@ -6,17 +6,25 @@ spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bi
 #endif // RTE16
 
 #include "types.comp"
-#include "generic_unary_head.comp"
 
-#if defined(DATA_A_IQ4_NL)
-// 16 invocations needed for init_iq4nl_shmem
-layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
+#if defined(SET_ROWS) && QUANT_K == 1
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+const uint BLOCK_SIZE = 512;
 #else
-layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+const uint BLOCK_SIZE = 32;
 #endif
 
 layout (binding = 0) readonly buffer S {float data_s[];};
+
+#if defined(SET_ROWS)
+#include "generic_binary_head.comp"
+layout (binding = 1) readonly buffer C {uvec2 data_i[];};
+layout (binding = 2) writeonly buffer Q {A_TYPE data_q[];};
+#else
+#include "generic_unary_head.comp"
 layout (binding = 1) writeonly buffer Q {A_TYPE data_q[];};
+#endif
 
 #if defined(DATA_A_Q4_0)
 void quantize(uint dst_idx, uint src_idx)
@@ -221,15 +229,56 @@ void quantize(uint dst_idx, uint src_idx)
 }
 #endif
 
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+void quantize(uint dst_idx, uint src_idx)
+{
+    data_q[dst_idx] = A_TYPE(data_s[src_idx]);
+}
+#endif
+
+#if defined(DATA_A_BF16)
+void quantize(uint dst_idx, uint src_idx)
+{
+    data_q[dst_idx] = A_TYPE(fp32_to_bf16(data_s[src_idx]));
+}
+#endif
+
+#if defined(SET_ROWS)
+
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
-    if (gl_LocalInvocationIndex.x != 0) {
-        return;
-    }
 #endif
 
-    const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K;
+    const uint idx = ((gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x) * BLOCK_SIZE + gl_LocalInvocationID.x) * QUANT_K;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    uint i00, i01, i02, i03;
+    get_indices(idx, i00, i01, i02, i03);
+
+    uint i12 = fastmod(i03, p.ne12);
+    uint i11 = fastmod(i02, p.ne11);
+    uint i10 = i01;
+
+    uint i1 = data_i[src1_idx(i10, i11, i12, 0) + get_boffset()].x;
+
+    uint src0_idx = src0_idx(i00, i01, i02, i03) + get_aoffset();
+    uint dst_idx = dst_idx(i00 / QUANT_K, i1, i02, i03) + get_doffset();
+
+    quantize(dst_idx, src0_idx);
+}
+
+#else
+
+void main() {
+#ifdef NEEDS_INIT_IQ_SHMEM
+    init_iq_shmem(gl_WorkGroupSize);
+#endif
+
+    const uint idx = (gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x) * QUANT_K;
 
     if (idx >= p.ne) {
         return;
@@ -240,3 +289,5 @@ void main() {
 
     quantize(dst_idx, src_idx);
 }
+
+#endif

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

@@ -18,6 +18,7 @@
 #extension GL_KHR_cooperative_matrix : enable
 #extension GL_KHR_memory_scope_semantics : enable
 #extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_ballot : enable
 #endif
 
 #ifdef MUL_MAT_ID
@@ -104,6 +105,10 @@ 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)
@@ -172,7 +177,47 @@ void main() {
     const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;
 
 #ifdef MUL_MAT_ID
-    uint _ne1 = 0;
+#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;
+    }
+
+    barrier();
+
+    _ne1 = _ne1_sh;
+#else
+    _ne1 = 0;
     for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
         for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
             if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
@@ -183,6 +228,7 @@ void main() {
     }
 
     barrier();
+#endif
 
     // Workgroup has no work
     if (ic * BN >= _ne1) return;

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

@@ -162,17 +162,32 @@ void main() {
         _ne1 = 0;
         uint num_elements = p.nei1 * p.nei0;
 
-        for (uint i = gl_SubgroupInvocationID; subgroupAny(i < num_elements); i += gl_SubgroupSize) {
+        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 ii0 = i % p.nei0;
             uint ii1 = i / p.nei0;
-            uint id = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+            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;
     }
@@ -414,17 +429,31 @@ void main() {
                 fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false);
             }
 
-            coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-            coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+            if ((ir + 1) * BM <= p.M && block_k + BK <= end_k) {
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
-            coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
+                coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID
-            coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
 #else
-            coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
 #endif
 
-            sum = coopMatMulAdd(mat_a, mat_b, sum);
+                sum = coopMatMulAdd(mat_a, mat_b, sum);
+            } else {
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+
+                coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
+#ifdef MUL_MAT_ID
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+#else
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
+#endif
+
+                sum = coopMatMulAdd(mat_a, mat_b, sum);
+            }
         }
 
         // Convert from ACC_TYPE to D_TYPE

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

@@ -0,0 +1,46 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+uint wrap_idx(int i, uint ne) {
+    if (i < 0) {
+        return i + ne;
+    } else if (i >= ne) {
+        return i - ne;
+    }
+    return i;
+}
+
+void main() {
+    const uint idx = get_idx();
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = fastdiv(idx, p.ne1_012mp, p.ne1_012L);
+    const uint i3_offset = i3 * p.ne12*p.ne11*p.ne10;
+    const uint i2 = fastdiv(idx - i3_offset, p.ne1_01mp, p.ne1_01L);
+    const uint i2_offset = i2*p.ne11*p.ne10;
+    const uint i1 = fastdiv(idx - i3_offset - i2_offset, p.ne1_0mp, p.ne1_0L);
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne10;
+
+    const uint p1 = floatBitsToUint(p.param1);
+    const uint p2 = floatBitsToUint(p.param2);
+    const int s0 = int(p1 >> 16)    - 0x8000;
+    const int s1 = int(p1 & 0xFFFF) - 0x8000;
+    const int s2 = int(p2 >> 16)    - 0x8000;
+    const int s3 = int(p2 & 0xFFFF) - 0x8000;
+
+    const uint i00 = wrap_idx(int(i0) - s0, p.ne10);
+    const uint i01 = wrap_idx(int(i1) - s1, p.ne11);
+    const uint i02 = wrap_idx(int(i2) - s2, p.ne12);
+    const uint i03 = wrap_idx(int(i3) - s3, p.ne13);
+
+    const uint a_idx = i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
+    const uint d_idx = i3 *p.nb13 + i2 *p.nb12 + i1 *p.nb11 + i0 *p.nb10;
+
+    data_d[get_doffset() + d_idx] = D_TYPE(data_a[get_aoffset() + a_idx]);
+}

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

@@ -3,6 +3,7 @@
 layout (push_constant) uniform parameter
 {
     uint ne; uint a_offset; uint d_offset;
+    uint ne00; uint ne01;
     uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13;
     float sf0; float sf1; float sf2; float sf3;
@@ -15,6 +16,61 @@ layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
+// from ggml.h: enum ggml_scale_mode, enum ggml_scale_flag
+#define NEAREST  0
+#define BILINEAR 1
+#define ALIGN_CORNERS (1 << 8)
+
+layout (constant_id = 0) const uint scale_mode = 0;
+
+float fetch_nearest(uint i10, uint i11, uint i12, uint i13) {
+    const uint i00 = uint(i10 / p.sf0);
+    const uint i01 = uint(i11 / p.sf1);
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+
+    return data_a[p.a_offset + i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00];
+}
+
+float fetch_bilinear(ivec2 c0, ivec2 c1, vec2 d, uint i12, uint i13) {
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+    const uint base = p.a_offset + i03 * p.nb03 + i02 * p.nb02;
+
+    const float v00 = data_a[base + c0.y * p.nb01 + c0.x * p.nb00];
+    const float v01 = data_a[base + c0.y * p.nb01 + c1.x * p.nb00];
+    const float v10 = data_a[base + c1.y * p.nb01 + c0.x * p.nb00];
+    const float v11 = data_a[base + c1.y * p.nb01 + c1.x * p.nb00];
+
+    return
+        v00 * (1.0-d.x) * (1.0-d.y) +
+        v01 * d.x       * (1.0-d.y) +
+        v10 * (1.0-d.x) * d.y +
+        v11 * d.x       * d.y;
+}
+
+float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) {
+    const ivec2 ne0 = ivec2(p.ne00, p.ne01);
+
+    const vec2 c = (vec2(i10, i11) + 0.5) / vec2(p.sf0, p.sf1) - 0.5;
+    const vec2 c0f = floor(c);
+    const vec2 d = c - c0f;
+    const ivec2 c0 = max(ivec2(c0f), 0);
+    const ivec2 c1 = min(ivec2(c0f + 1), ne0 - 1);
+
+    return fetch_bilinear(c0, c1, d, i12, i13);
+}
+
+float interpolate_bilinear_align_corners(uint i10, uint i11, uint i12, uint i13) {
+    const vec2 c = vec2(i10, i11) / vec2(p.sf0, p.sf1);
+    const vec2 c0f = floor(c);
+    const vec2 d = c - c0f;
+    const ivec2 c0 = ivec2(c0f);
+    const ivec2 c1 = c0 + 1;
+
+    return fetch_bilinear(c0, c1, d, i12, i13);
+}
+
 void main() {
     const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 
@@ -27,10 +83,18 @@ void main() {
     const uint i12 = (idx / (p.ne10 * p.ne11)) % p.ne12;
     const uint i13 = (idx / (p.ne10 * p.ne11 * p.ne12)) % p.ne13;
 
-    const uint i00 = uint(i10 / p.sf0);
-    const uint i01 = uint(i11 / p.sf1);
-    const uint i02 = uint(i12 / p.sf2);
-    const uint i03 = uint(i13 / p.sf3);
+    float result;
+    switch (scale_mode) {
+        case NEAREST:
+            result = fetch_nearest(i10, i11, i12, i13);
+            break;
+        case BILINEAR:
+            result = interpolate_bilinear(i10, i11, i12, i13);
+            break;
+        case BILINEAR | ALIGN_CORNERS:
+            result = interpolate_bilinear_align_corners(i10, i11, i12, i13);
+            break;
+    }
 
-    data_d[p.d_offset + idx] = D_TYPE(data_a[p.a_offset + i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]);
+    data_d[p.d_offset + idx] = D_TYPE(result);
 }

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

@@ -518,6 +518,11 @@ void process_shaders() {
         string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }
 
+    for (std::string t : {"f32", "f16", "bf16", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
+        string_to_spv("set_rows_" + t, "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+        string_to_spv("set_rows_" + t + "_rte", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
+    }
+
     auto get_type_str = [](bool f16) {
         return f16 ? "float16_t" : "float";
     };
@@ -648,6 +653,8 @@ 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("roll_f32", "roll.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+
     for (auto &c : compiles) {
         c.wait();
     }

gguf-py/gguf/constants.py

@@ -187,6 +187,9 @@ class Keys:
     class Classifier:
         OUTPUT_LABELS = "{arch}.classifier.output_labels"
 
+    class ShortConv:
+        L_CACHE = "{arch}.shortconv.l_cache"
+
     class Tokenizer:
         MODEL                = "tokenizer.ggml.model"
         PRE                  = "tokenizer.ggml.pre"
@@ -352,6 +355,7 @@ class MODEL_ARCH(IntEnum):
     EXAONE           = auto()
     GRANITE          = auto()
     GRANITE_MOE      = auto()
+    GRANITE_HYBRID   = auto()
     CHAMELEON        = auto()
     WAVTOKENIZER_DEC = auto()
     PLM              = auto()
@@ -361,6 +365,7 @@ class MODEL_ARCH(IntEnum):
     ERNIE4_5         = auto()
     HUNYUAN_MOE      = auto()
     SMOLLM3          = auto()
+    LFM2             = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -532,6 +537,9 @@ class MODEL_TENSOR(IntEnum):
     POSNET_ATTN_K        = auto()
     POSNET_ATTN_V        = auto()
     POSNET_ATTN_OUT      = auto()
+    SHORTCONV_CONV       = auto()
+    SHORTCONV_INPROJ     = auto()
+    SHORTCONV_OUTPROJ    = auto()
     # vision
     V_MMPROJ             = auto()
     V_MMPROJ_FC          = auto()
@@ -661,6 +669,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.EXAONE:           "exaone",
     MODEL_ARCH.GRANITE:          "granite",
     MODEL_ARCH.GRANITE_MOE:      "granitemoe",
+    MODEL_ARCH.GRANITE_HYBRID:   "granitehybrid",
     MODEL_ARCH.CHAMELEON:        "chameleon",
     MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
     MODEL_ARCH.PLM:              "plm",
@@ -671,6 +680,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.FALCON_H1:        "falcon-h1",
     MODEL_ARCH.HUNYUAN_MOE:      "hunyuan-moe",
     MODEL_ARCH.SMOLLM3:          "smollm3",
+    MODEL_ARCH.LFM2:             "lfm2",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -842,6 +852,9 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.POSNET_ATTN_K:             "posnet.{bid}.attn_k",
     MODEL_TENSOR.POSNET_ATTN_V:             "posnet.{bid}.attn_v",
     MODEL_TENSOR.POSNET_ATTN_OUT:           "posnet.{bid}.attn_output",
+    MODEL_TENSOR.SHORTCONV_CONV:            "blk.{bid}.shortconv.conv",
+    MODEL_TENSOR.SHORTCONV_INPROJ:          "blk.{bid}.shortconv.in_proj",
+    MODEL_TENSOR.SHORTCONV_OUTPROJ:         "blk.{bid}.shortconv.out_proj",
     # vision
     MODEL_TENSOR.V_MMPROJ:                  "mm.{bid}",
     MODEL_TENSOR.V_MMPROJ_FC:               "mm.model.fc",
@@ -2143,6 +2156,36 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP_SHEXP,
         MODEL_TENSOR.FFN_DOWN_SHEXP,
     ],
+    MODEL_ARCH.GRANITE_HYBRID: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.SSM_IN,
+        MODEL_TENSOR.SSM_CONV1D,
+        MODEL_TENSOR.SSM_DT,
+        MODEL_TENSOR.SSM_A,
+        MODEL_TENSOR.SSM_D,
+        MODEL_TENSOR.SSM_NORM,
+        MODEL_TENSOR.SSM_OUT,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        # MoE
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        # Dense
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     MODEL_ARCH.CHAMELEON: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -2324,6 +2367,24 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.LFM2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.SHORTCONV_CONV,
+        MODEL_TENSOR.SHORTCONV_INPROJ,
+        MODEL_TENSOR.SHORTCONV_OUTPROJ,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.ATTN_NORM, # operator_norm
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+    ],
     # TODO
 }
 

gguf-py/gguf/gguf_writer.py

@@ -648,6 +648,9 @@ class GGUFWriter:
     def add_convnext_block_count(self, length: int) -> None:
         self.add_uint32(Keys.ConvNext.BLOCK_COUNT.format(arch=self.arch), length)
 
+    def add_shortconv_l_cache(self, length: int) -> None:
+        self.add_uint32(Keys.ShortConv.L_CACHE.format(arch=self.arch), length)
+
     def add_block_count(self, length: int) -> None:
         self.add_uint32(Keys.LLM.BLOCK_COUNT.format(arch=self.arch), length)
 

gguf-py/gguf/tensor_mapping.py

@@ -13,7 +13,7 @@ class TensorNameMap:
             "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx jais exaone
             "transformer.word_embeddings",               # falcon
             "word_embeddings",                           # bloom
-            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo2 rwkv6qwen2 glm4-0414
+            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo2 rwkv6qwen2 glm4-0414 granite-hybrid
             "tok_embeddings",                            # llama-pth
             "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
@@ -50,6 +50,7 @@ class TensorNameMap:
             "model.pre_ln",               # rwkv7
             "model.layers.0.pre_norm",    # rwkv7
             "backbone.norm",              # wavtokenizer
+            "model.embedding_norm",       # lfm2
         ),
 
         # Position embeddings
@@ -118,7 +119,7 @@ class TensorNameMap:
             "transformer.h.{bid}.input_layernorm",                  # falcon7b
             "h.{bid}.input_layernorm",                              # bloom
             "transformer.h.{bid}.ln_mlp",                           # falcon40b
-            "model.layers.{bid}.input_layernorm",                   # llama-hf nemotron olmoe phimoe
+            "model.layers.{bid}.input_layernorm",                   # llama-hf nemotron olmoe phimoe granite-hybrid
             "layers.{bid}.attention_norm",                          # llama-pth
             "language_model.encoder.layers.{bid}.input_layernorm",  # persimmon
             "model.layers.{bid}.ln1",                               # yi
@@ -136,6 +137,7 @@ class TensorNameMap:
             "model.layers.{bid}.ln1",                               # rwkv7
             "model.layers.{bid}.input_layernorm",                   # llama4
             "transformer_encoder.{bid}.attention_norm",             # neobert
+            "model.layers.{bid}.operator_norm",                     # lfm2
         ),
 
         # Attention norm 2
@@ -220,6 +222,7 @@ class TensorNameMap:
             "transformer.h.{bid}.self_attention.dense",                     # falcon
             "h.{bid}.self_attention.dense",                                 # bloom
             "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe olmo2 phimoe
+            "model.layers.{bid}.self_attn.out_proj",                        # lfm2
             "model.layers.{bid}.self_attn.linear_attn",                     # deci
             "layers.{bid}.attention.wo",                                    # llama-pth
             "encoder.layer.{bid}.attention.output.dense",                   # bert
@@ -279,7 +282,7 @@ class TensorNameMap:
             "transformer.decoder_layer.{bid}.rms_norm_2",                    # Grok
             "encoder.layers.{bid}.post_attention_layernorm",                 # chatglm
             "transformer.layers.{bid}.ffn_norm",                             # openelm
-            "model.layers.{bid}.pre_ff_layernorm",                           # jamba
+            "model.layers.{bid}.pre_ff_layernorm",                           # jamba granite-hybrid
             "model.layers.{bid}.pre_moe_layernorm",                          # mini-jamba
             "model.layers.{bid}.post_attention_layernorm",                   # llama4
             "transformer_encoder.{bid}.ffn_norm",                            # neobert
@@ -349,7 +352,7 @@ class TensorNameMap:
             "model.layers.{bid}.residual_mlp.w3",                     # arctic
             "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
             "transformer.h.{bid}.mlp.c_fc_1",                         # exaone
-            "model.layers.{bid}.feed_forward.up_proj",                # llama4 jamba
+            "model.layers.{bid}.feed_forward.up_proj",                # llama4 jamba granite-hybrid
             "transformer_encoder.{bid}.ffn.w12",                      # neobert
         ),
 
@@ -389,7 +392,7 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.linear_1",           # refact
             "model.layers.{bid}.residual_mlp.w1",         # arctic
             "transformer.h.{bid}.mlp.c_fc_0",             # exaone
-            "model.layers.{bid}.feed_forward.gate_proj",  # llama4 jamba
+            "model.layers.{bid}.feed_forward.gate_proj",  # llama4 jamba granite-hybrid
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
@@ -435,7 +438,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.mlp.down_layer",                     # jina-bert-v2
             "encoder.layers.{bid}.mlp.dense_4h_to_h",                 # chatglm
             "model.layers.h.{bid}.mlp.c_proj",                        # exaone
-            "model.layers.{bid}.feed_forward.down_proj",              # llama4 jamba
+            "model.layers.{bid}.feed_forward.down_proj",              # llama4 jamba granite-hybrid
             "transformer_encoder.{bid}.ffn.w3",                       # neobert
         ),
 
@@ -558,13 +561,13 @@ class TensorNameMap:
         MODEL_TENSOR.SSM_IN: (
             "model.layers.{bid}.in_proj",           # mamba-hf
             "backbone.layers.{bid}.mixer.in_proj",  # mamba
-            "model.layers.{bid}.mamba.in_proj",     # jamba falcon-h1
+            "model.layers.{bid}.mamba.in_proj",     # jamba falcon-h1 granite-hybrid
         ),
 
         MODEL_TENSOR.SSM_CONV1D: (
             "model.layers.{bid}.conv1d",           # mamba-hf
             "backbone.layers.{bid}.mixer.conv1d",  # mamba
-            "model.layers.{bid}.mamba.conv1d",     # jamba falcon-h1
+            "model.layers.{bid}.mamba.conv1d",     # jamba falcon-h1 granite-hybrid
         ),
 
         MODEL_TENSOR.SSM_X: (
@@ -576,7 +579,7 @@ class TensorNameMap:
         MODEL_TENSOR.SSM_DT: (
             "model.layers.{bid}.dt_proj",           # mamba-hf
             "backbone.layers.{bid}.mixer.dt_proj",  # mamba
-            "model.layers.{bid}.mamba.dt_proj",     # jamba falcon-h1
+            "model.layers.{bid}.mamba.dt_proj",     # jamba falcon-h1 granite-hybrid
         ),
 
         MODEL_TENSOR.SSM_DT_NORM: (
@@ -586,7 +589,7 @@ class TensorNameMap:
         MODEL_TENSOR.SSM_A: (
             "model.layers.{bid}.A_log",           # mamba-hf
             "backbone.layers.{bid}.mixer.A_log",  # mamba
-            "model.layers.{bid}.mamba.A_log",     # jamba falcon-h1
+            "model.layers.{bid}.mamba.A_log",     # jamba falcon-h1 granite-hybrid
         ),
 
         MODEL_TENSOR.SSM_B_NORM: (
@@ -602,18 +605,18 @@ class TensorNameMap:
         MODEL_TENSOR.SSM_D: (
             "model.layers.{bid}.D",           # mamba-hf
             "backbone.layers.{bid}.mixer.D",  # mamba
-            "model.layers.{bid}.mamba.D",     # jamba falcon-h1
+            "model.layers.{bid}.mamba.D",     # jamba falcon-h1 granite-hybrid
         ),
 
         MODEL_TENSOR.SSM_NORM: (
-            "model.layers.{bid}.mamba.norm", # falcon-h1
+            "model.layers.{bid}.mamba.norm", # falcon-h1 granite-hybrid
             "backbone.layers.{bid}.mixer.norm",  # mamba2
         ),
 
         MODEL_TENSOR.SSM_OUT: (
             "model.layers.{bid}.out_proj",           # mamba-hf
             "backbone.layers.{bid}.mixer.out_proj",  # mamba
-            "model.layers.{bid}.mamba.out_proj",     # jamba falcon-h1
+            "model.layers.{bid}.mamba.out_proj",     # jamba falcon-h1 granite-hybrid
         ),
 
         MODEL_TENSOR.TIME_MIX_W0: (
@@ -1015,6 +1018,18 @@ class TensorNameMap:
             "backbone.posnet.{bid}.proj_out", # wavtokenizer
         ),
 
+        MODEL_TENSOR.SHORTCONV_CONV: (
+            "model.layers.{bid}.conv.conv",
+        ),
+
+        MODEL_TENSOR.SHORTCONV_INPROJ: (
+            "model.layers.{bid}.conv.in_proj",
+        ),
+
+        MODEL_TENSOR.SHORTCONV_OUTPROJ: (
+            "model.layers.{bid}.conv.out_proj",
+        ),
+
         #############################################################################
         ## Vision encoder
 

include/llama.h

@@ -79,47 +79,6 @@ extern "C" {
         LLAMA_VOCAB_TYPE_RWKV = 5, // RWKV tokenizer based on greedy tokenization
     };
 
-    // pre-tokenization types
-    enum llama_vocab_pre_type {
-        LLAMA_VOCAB_PRE_TYPE_DEFAULT        = 0,
-        LLAMA_VOCAB_PRE_TYPE_LLAMA3         = 1,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM   = 2,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,
-        LLAMA_VOCAB_PRE_TYPE_FALCON         = 4,
-        LLAMA_VOCAB_PRE_TYPE_MPT            = 5,
-        LLAMA_VOCAB_PRE_TYPE_STARCODER      = 6,
-        LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
-        LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
-        LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
-        LLAMA_VOCAB_PRE_TYPE_STABLELM2      = 10,
-        LLAMA_VOCAB_PRE_TYPE_QWEN2          = 11,
-        LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,
-        LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
-        LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,
-        LLAMA_VOCAB_PRE_TYPE_PORO           = 15,
-        LLAMA_VOCAB_PRE_TYPE_CHATGLM3       = 16,
-        LLAMA_VOCAB_PRE_TYPE_CHATGLM4       = 17,
-        LLAMA_VOCAB_PRE_TYPE_VIKING         = 18,
-        LLAMA_VOCAB_PRE_TYPE_JAIS           = 19,
-        LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
-        LLAMA_VOCAB_PRE_TYPE_SMOLLM         = 21,
-        LLAMA_VOCAB_PRE_TYPE_CODESHELL      = 22,
-        LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
-        LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
-        LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
-        LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
-        LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 28,
-        LLAMA_VOCAB_PRE_TYPE_GPT4O          = 29,
-        LLAMA_VOCAB_PRE_TYPE_SUPERBPE       = 30,
-        LLAMA_VOCAB_PRE_TYPE_TRILLION       = 31,
-        LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
-        LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
-        LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
-        LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
-        LLAMA_VOCAB_PRE_TYPE_HUNYUAN        = 36,
-    };
-
     enum llama_rope_type {
         LLAMA_ROPE_TYPE_NONE   = -1,
         LLAMA_ROPE_TYPE_NORM   = 0,

models/templates/llama-cpp-rwkv-world.jinja

@@ -0,0 +1,34 @@
+{%- if not add_generation_prompt is defined -%}
+    {%- set add_generation_prompt = true -%}
+{%- endif -%}
+{%- set ns = namespace(system_prompt='') -%}
+{%- for message in messages -%}
+    {%- if message['role'] == 'system' -%}
+        {%- set ns.system_prompt = message['content'] -%}
+    {%- endif -%}
+{%- endfor -%}
+{{bos_token}}
+{%- if ns.system_prompt != '' -%}
+{{- 'System: ' + ns.system_prompt + '\n\n' -}}
+{%- endif -%}
+{%- for message in messages -%}
+    {%- if message['role'] == 'user' -%}
+        {{- 'User: ' + message['content']|trim + '\n\n' -}}
+    {%- endif -%}
+    {%- if message['role'] == 'assistant' and message['content'] is  not none -%}
+        {%- set content = message['content'] -%}
+        {%- if '</think>' in content -%}
+            {%- set content = content.split('</think>')[-1] -%}
+        {%- endif -%}
+        {{- 'Assistant: ' + content|trim + '\n\n' -}}
+    {%- endif -%}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+    {{- 'Assistant:' -}}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- ' <think>\n</think>' }}
+    {%- endif %}
+    {%- if enable_thinking is defined and enable_thinking is true %}
+        {{- ' <think>' }}
+    {%- endif %}
+{%- endif -%}

requirements/requirements-all.txt

@@ -3,6 +3,7 @@
 -r ../tools/server/tests/requirements.txt
 
 -r ./requirements-compare-llama-bench.txt
+-r ./requirements-server-bench.txt
 -r ./requirements-pydantic.txt
 -r ./requirements-test-tokenizer-random.txt
 

requirements/requirements-server-bench.txt

@@ -0,0 +1,5 @@
+datasets~=3.2.0
+matplotlib~=3.10.0
+numpy~=1.26.4
+requests~=2.32.3
+tqdm~=4.67.1

scripts/create_ops_docs.py

@@ -0,0 +1,196 @@
+#!/usr/bin/env python3
+
+"""
+This script parses docs/ops/*.csv and creates the ops.md, which is a table documenting supported operations on various ggml backends.
+"""
+import csv
+import logging
+import sys
+from pathlib import Path
+from collections import defaultdict
+
+
+class DocsGenerator:
+    def __init__(self, ggml_root: str, output_filename: str = "ops.md"):
+        self.ggml_root = Path(ggml_root)
+        self.ops_dir = self.ggml_root / "docs" / "ops"
+        self.output_filename = output_filename
+        self.backend_support: dict[str, dict[str, list[bool]]] = defaultdict(
+            lambda: defaultdict(list)
+        )
+        self.all_operations: set[str] = set()
+        self.all_backends: set[str] = set()
+        self.logger = logging.getLogger(__name__)
+
+    def parse_support_files(self) -> None:
+        if not self.ops_dir.exists():
+            self.logger.warning(f"ops directory not found: {self.ops_dir}")
+            return
+
+        self.logger.info(f"Parsing support files from {self.ops_dir}...")
+
+        for support_file in self.ops_dir.glob("*.csv"):
+            self.logger.info(f"  Reading: {support_file.name}")
+            self._parse_support_file(support_file)
+
+    def _parse_support_file(self, file_path: Path) -> None:
+        try:
+            with open(file_path, "r", newline='') as f:
+                reader = csv.DictReader(f)
+
+                for row in reader:
+                    # Skip rows that don't have support mode
+                    if row.get('test_mode') != 'support':
+                        continue
+
+                    backend_name = row.get('backend_name', '').strip()
+                    operation = row.get('op_name', '').strip()
+                    supported_str = row.get('error_message', '').strip()  # "yes" or "no"
+                    backend_reg_name = row.get('backend_reg_name', '').strip()
+
+                    # Skip invalid or error operations
+                    if not operation or not backend_name or operation in [
+                        "CONTEXT_ERROR",
+                        "BUILD_ERROR",
+                    ]:
+                        continue
+
+                    is_supported = supported_str.lower() == "yes"
+
+                    # Use backend_reg_name for grouping, fallback to backend_name
+                    backend_key = backend_reg_name if backend_reg_name else backend_name
+
+                    self.all_backends.add(backend_key)
+                    self.backend_support[backend_key][operation].append(is_supported)
+                    self.all_operations.add(operation)
+
+        except Exception as e:
+            self.logger.error(f"    Error parsing {file_path}: {e}")
+
+    def get_backend_support_status(self, backend: str, operation: str) -> str:
+        support_list = self.backend_support[backend].get(operation, [])
+
+        if not support_list:
+            return "unsupported"
+
+        all_supported = all(support_list)
+        any_supported = any(support_list)
+
+        if all_supported:
+            return "supported"
+        elif any_supported:
+            return "partially supported"
+        else:
+            return "unsupported"
+
+    def get_support_status(self, operation: str) -> str:
+        if operation not in self.all_operations:
+            return "unsupported"
+
+        support_count = 0
+        total_backends = len(self.all_backends)
+
+        for backend in self.all_backends:
+            if self.backend_support[backend].get(operation, False):
+                support_count += 1
+
+        if support_count == 0:
+            return "unsupported"
+        elif support_count == total_backends:
+            return "supported"
+        else:
+            return "partially supported"
+
+    def get_support_symbol(self, status: str) -> str:
+        symbols = {"supported": "✅", "partially supported": "🟡", "unsupported": "❌"}
+        return symbols.get(status, "❓")
+
+    def generate_markdown(self) -> str:
+        lines = []
+
+        lines.append("# GGML Operations")
+        lines.append("")
+        lines.append("List of GGML operations and backend support status.")
+        lines.append("")
+        lines.append("Legend:")
+        lines.append("- ✅ Fully supported by this backend")
+        lines.append("- 🟡 Partially supported by this backend")
+        lines.append("- ❌ Not supported by this backend")
+        lines.append("")
+
+        backends = sorted(self.all_backends)
+        header = "| Operation |"
+        for backend in backends:
+            header += f" {backend} |"
+
+        separator = "|-----------|"
+        for _ in backends:
+            separator += "------|"
+
+        lines.append(header)
+        lines.append(separator)
+
+        sorted_operations = sorted(self.all_operations)
+
+        for operation in sorted_operations:
+            row = f"| {operation:>32} |"
+
+            for backend in backends:
+                status = self.get_backend_support_status(backend, operation)
+                if status == "supported":
+                    symbol = "✅"
+                elif status == "partially supported":
+                    symbol = "🟡"
+                else:
+                    symbol = "❌"
+                row += f" {symbol} |"
+
+            lines.append(row)
+
+        lines.append("")
+
+        return "\n".join(lines)
+
+    def run(self) -> None:
+        self.logger.info("Parsing GGML operation support files...")
+        self.parse_support_files()
+
+        if not self.all_operations:
+            self.logger.error(
+                "No operations found. Make sure to run test-backend-ops support --output csv > docs/ops/file.csv first."
+            )
+            return
+
+        self.logger.info(
+            f"Found {len(self.all_operations)} operations across {len(self.all_backends)} backends"
+        )
+
+        self.logger.info("Generating markdown...")
+        markdown_content = self.generate_markdown()
+
+        docs_dir = self.ggml_root / "docs"
+        docs_dir.mkdir(exist_ok=True)
+
+        ops_file = docs_dir / self.output_filename
+        with open(ops_file, "w") as f:
+            f.write(markdown_content)
+
+        self.logger.info(f"Generated: {ops_file}")
+        self.logger.info(f"Operations: {len(self.all_operations)}")
+        self.logger.info(f"Backends: {len(self.all_backends)}")
+
+
+def main():
+    logging.basicConfig(level=logging.INFO)
+
+    if len(sys.argv) > 1:
+        output_filename = sys.argv[1]
+    else:
+        output_filename = "ops.md"
+
+    generator = DocsGenerator(".", output_filename)
+    generator.run()
+
+
+if __name__ == "__main__":
+    main()

scripts/server-bench.py

@@ -0,0 +1,210 @@
+#!/usr/bin/env python3
+
+import argparse
+import json
+import subprocess
+from time import sleep, time
+from typing import Optional
+
+import datasets
+import logging
+import matplotlib.pyplot as plt
+import numpy as np
+import requests
+from tqdm.contrib.concurrent import thread_map
+
+
+logging.basicConfig(level=logging.INFO, format='%(message)s')
+logger = logging.getLogger("server-bench")
+
+
+def get_prompts(n_prompts: int) -> list[str]:
+    logger.info("Loading MMLU dataset...")
+    ret = datasets.load_dataset("cais/mmlu", "all")["test"]["question"]  # type: ignore
+    if n_prompts >= 0:
+        ret = ret[:n_prompts]
+    return ret
+
+
+def get_server(path_server: str, path_model: str, path_log: Optional[str], port: int, n_gpu_layers: int, parallel: int, ctx_size: int) -> dict:
+    logger.info("Starting the llama.cpp server...")
+    address = f"http://localhost:{port}"
+
+    popen_args: list[str] = [
+        path_server,
+        "--flash-attn",
+        "--n-gpu-layers", str(n_gpu_layers),
+        "--parallel", str(parallel),
+        "--ctx-size", str(parallel * ctx_size),
+        "--model", path_model,
+        "--port", str(port),
+        "--swa-full",  # FIXME performance bad otherwise
+        # "--attn-streams",
+    ]
+    fout = open("bench.log", "w") if path_log is not None else subprocess.DEVNULL
+    process = subprocess.Popen(popen_args, stdout=fout, stderr=subprocess.STDOUT)
+
+    n_failures: int = 0
+    while True:
+        try:
+            sleep(1.0)
+            exit_code = process.poll()
+            if exit_code is not None:
+                raise RuntimeError(f"llama.cpp server for {path_model} exited unexpectedly with exit code {exit_code}")
+            response = requests.get(f"{address}/health")
+            if response.status_code == 200:
+                break
+        except requests.ConnectionError:
+            n_failures += 1
+            if n_failures >= 10:
+                raise RuntimeError(f"llama.cpp server for {path_model} is not healthy after 10 seconds")
+
+    return {"process": process, "address": address, "fout": fout}
+
+
+def get_prompt_length(data: dict) -> int:
+    session = data["session"]
+    server_address: str = data["server_address"]
+
+    response = session.post(
+        f"{server_address}/apply-template",
+        json={"messages": [{"role": "user", "content": data["prompt"], "stream": True}]}
+    )
+    if response.status_code != 200:
+        raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
+    prompt: str = json.loads(response.text)["prompt"]
+    response = session.post(
+        f"{server_address}/tokenize",
+        json={"content": prompt, "add_special": True}
+    )
+    if response.status_code != 200:
+        raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
+    tokens: list[str] = json.loads(response.text)["tokens"]
+    return len(tokens)
+
+
+def send_prompt(data: dict) -> tuple[float, list[float]]:
+    session = data["session"]
+    server_address: str = data["server_address"]
+
+    response = session.post(
+        f"{server_address}/apply-template",
+        json={"messages": [{"role": "user", "content": data["prompt"], "stream": True}]}
+    )
+    if response.status_code != 200:
+        raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
+    prompt: str = json.loads(response.text)["prompt"]
+
+    json_data: dict = {"prompt": prompt, "seed": data["seed"], "n_predict": data["n_predict"], "stream": True}
+    response = session.post(f"{server_address}/completion", json=json_data, stream=True)
+
+    last_valid_line: str = ""
+    token_arrival_times: list[float] = []
+    for line in response.iter_lines(decode_unicode=True):
+        if not line.startswith("data: "):
+            continue
+        last_valid_line = line
+        token_arrival_times.append(time())
+    token_arrival_times = token_arrival_times[:-1]
+
+    if response.status_code != 200:
+        raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
+    timings: dict = json.loads(last_valid_line[6:])["timings"]
+
+    return (timings["prompt_ms"], token_arrival_times)
+
+
+def benchmark(path_server: str, path_model: str, path_log: Optional[str], port: int, n_gpu_layers: int, parallel: int, ctx_size: int, n_prompts: int, n_predict: int):
+    num_workers: int = parallel + 1
+    prompts: list[str] = get_prompts(n_prompts)
+
+    server: Optional[dict] = None
+    session = None
+    try:
+        server = get_server(path_server, path_model, path_log, port, n_gpu_layers, parallel, ctx_size)
+        server_address: str = server["address"]
+
+        adapter = requests.adapters.HTTPAdapter(pool_connections=num_workers, pool_maxsize=num_workers)  # type: ignore
+        session = requests.Session()
+        session.mount("http://", adapter)
+        session.mount("https://", adapter)
+
+        data: list[dict] = []
+        for i, p in enumerate(prompts):
+            data.append({"session": session, "server_address": server_address, "prompt": p, "n_predict": n_predict, "seed": i})
+
+        logger.info("Getting the prompt lengths...")
+        prompt_n = [get_prompt_length(d) for d in data]
+
+        logger.info("Starting the benchmark...\n")
+        t0 = time()
+        results: list[tuple[int, list[float]]] = thread_map(send_prompt, data, max_workers=num_workers, chunksize=1)
+    finally:
+        if server is not None:
+            server["process"].terminate()
+            server["process"].wait()
+        if session is not None:
+            session.close()
+
+    prompt_ms = []
+    token_t = []
+    depth_sum: int = 0
+    for pn, (pms, tat) in zip(prompt_n, results):
+        prompt_ms.append(pms)
+        token_t += tat
+        n_tokens: int = len(tat)
+        depth_sum += n_tokens * pn
+        depth_sum += n_tokens * (n_tokens + 1) // 2
+    prompt_n = np.array(prompt_n, dtype=np.int64)
+    prompt_ms = np.array(prompt_ms, dtype=np.float64)
+    token_t = np.array(token_t, dtype=np.float64)
+
+    token_t -= t0
+    token_t_last = np.max(token_t)
+
+    logger.info("")
+    logger.info(f"Benchmark duration:                {token_t_last:.2f} s")
+    logger.info(f"Request throughput:                {n_prompts / token_t_last:.2f} requests/s = {n_prompts / (token_t_last/60):.2f} requests/min")
+    logger.info(f"Total prompt length:               {np.sum(prompt_n)} tokens")
+    logger.info(f"Average prompt length:             {np.mean(prompt_n):.2f} tokens")
+    logger.info(f"Average prompt latency:            {np.mean(prompt_ms):.2f} ms")
+    logger.info(f"Average prompt speed:              {np.sum(prompt_n) / (1e-3 * np.sum(prompt_ms)):.2f} tokens/s")
+    logger.info(f"Total generated tokens:            {token_t.shape[0]}")
+    logger.info(f"Average generation depth:          {depth_sum / token_t.shape[0]:.2f} tokens")
+    logger.info(f"Average total generation speed:    {token_t.shape[0] / token_t_last:.2f} tokens/s")
+    logger.info(f"Average generation speed per slot: {token_t.shape[0] / (parallel * token_t_last):.2f} tokens/s / slot")
+
+    plt.figure()
+    plt.scatter(prompt_n, prompt_ms, s=10.0, marker=".", alpha=0.25)
+    plt.xlim(0, 1.05 * np.max(prompt_n))
+    plt.ylim(0, 1.05 * np.max(prompt_ms))
+    plt.title(path_model)
+    plt.xlabel("Prompt length [tokens]")
+    plt.ylabel("Time to first token [ms]")
+    plt.savefig("prompt_time.png", dpi=240)
+
+    bin_max = np.ceil(token_t_last) + 1
+    plt.figure()
+    plt.hist(token_t, np.arange(0, bin_max))
+    plt.xlim(0, bin_max + 1)
+    plt.title(path_model)
+    plt.xlabel("Time [s]")
+    plt.ylabel("Num. tokens generated per second")
+    plt.savefig("gen_rate.png", dpi=240)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Tool for benchmarking the throughput of the llama.cpp HTTP server. "
+        "Results are printed to console and visualized as plots (saved to current working directory).")
+    parser.add_argument("--path_server", type=str, default="llama-server", help="Path to the llama.cpp server binary")
+    parser.add_argument("--path_model", type=str, required=True, help="Path to the model to use for the benchmark")
+    parser.add_argument("--path_log", type=str, default=None, help="Path to the model to use for the benchmark")
+    parser.add_argument("--port", type=int, default=18725, help="Port to use for the server during the benchmark")
+    parser.add_argument("--n_gpu_layers", type=int, default=999, help="Number of GPU layers for the server")
+    parser.add_argument("--parallel", type=int, default=16, help="Number of slots for the server")
+    parser.add_argument("--ctx_size", type=int, default=4096, help="Server context size per slot")
+    parser.add_argument("--n_prompts", type=int, default=1000, help="Number of prompts to evaluate")
+    parser.add_argument("--n_predict", type=int, default=2048, help="Max. number of tokens to predict per prompt")
+    args = parser.parse_args()
+    benchmark(**vars(args))

scripts/sync-ggml.last

@@ -1 +1 @@
-0405219965324e11a29b6aadfe22a6d66131978f
+d62df60a07ba3deeb85e5cfc9b1ee07645ff35e2

src/llama-arch.cpp

@@ -73,6 +73,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_ARWKV7,           "arwkv7"           },
     { LLM_ARCH_GRANITE,          "granite"          },
     { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
+    { LLM_ARCH_GRANITE_HYBRID,   "granitehybrid"    },
     { LLM_ARCH_CHAMELEON,        "chameleon"        },
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_PLM,              "plm"              },
@@ -82,6 +83,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
     { LLM_ARCH_HUNYUAN_MOE,      "hunyuan-moe"      },
     { LLM_ARCH_SMOLLM3,          "smollm3"          },
+    { LLM_ARCH_LFM2,             "lfm2"             },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -154,7 +156,6 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
-    { LLM_KV_ATTENTION_LAYER_INDICES,                "%s.attention.layer_indices"                },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -188,6 +189,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
 
     { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
 
+    { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" },
+
     { LLM_KV_TOKENIZER_MODEL,                "tokenizer.ggml.model"                    },
     { LLM_KV_TOKENIZER_PRE,                  "tokenizer.ggml.pre"                      },
     { LLM_KV_TOKENIZER_LIST,                 "tokenizer.ggml.tokens"                   },
@@ -1641,6 +1644,43 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_GRANITE_HYBRID,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
+            { LLM_TENSOR_OUTPUT,         "output" },
+            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
+            // mamba(2) ssm layers
+            { LLM_TENSOR_SSM_IN,         "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,     "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,         "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,          "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,          "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,       "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,        "blk.%d.ssm_out" },
+            // attention layers
+            { LLM_TENSOR_ATTN_Q,         "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,         "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,         "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,       "blk.%d.attn_output" },
+            // dense FFN
+            { LLM_TENSOR_FFN_NORM,       "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,       "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
+            // moe FFN
+            { LLM_TENSOR_FFN_NORM,       "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,   "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,  "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,  "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,    "blk.%d.ffn_up_exps" },
+            // shared expert
+            { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,   "blk.%d.ffn_up_shexp" },
+        },
+    },
     {
         LLM_ARCH_CHAMELEON,
         {
@@ -1777,26 +1817,47 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
         },
     },
     {
-        LLM_ARCH_UNKNOWN,
+        LLM_ARCH_SMOLLM3,
         {
-            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
+            { LLM_TENSOR_OUTPUT,         "output" },
+            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,         "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,         "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,         "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,       "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,       "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,       "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
         },
     },
     {
-        LLM_ARCH_SMOLLM3,
+        LLM_ARCH_LFM2,
         {
-            { LLM_TENSOR_TOKEN_EMBD,     "token_embd"            },
-            { LLM_TENSOR_OUTPUT_NORM,    "output_norm"           },
-            { LLM_TENSOR_OUTPUT,         "output"                },
-            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm"      },
-            { LLM_TENSOR_ATTN_Q,         "blk.%d.attn_q"         },
-            { LLM_TENSOR_ATTN_K,         "blk.%d.attn_k"         },
-            { LLM_TENSOR_ATTN_V,         "blk.%d.attn_v"         },
-            { LLM_TENSOR_ATTN_OUT,       "blk.%d.attn_output"    },
-            { LLM_TENSOR_FFN_NORM,       "blk.%d.ffn_norm"       },
-            { LLM_TENSOR_FFN_GATE,       "blk.%d.ffn_gate"       },
-            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down"       },
-            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up"         },
+            { LLM_TENSOR_ATTN_NORM,         "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,            "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,            "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,            "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,          "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_K_NORM,       "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_Q_NORM,       "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_FFN_DOWN,          "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_GATE,          "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_NORM,          "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,            "blk.%d.ffn_up" },
+            { LLM_TENSOR_SHORTCONV_CONV,    "blk.%d.shortconv.conv" },
+            { LLM_TENSOR_SHORTCONV_INPROJ,  "blk.%d.shortconv.in_proj" },
+            { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
+            { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+        }
+    },
+    {
+        LLM_ARCH_UNKNOWN,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
         },
     },
 };
@@ -1960,6 +2021,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_CONVNEXT_PW1,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CONVNEXT_PW2,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CONVNEXT_GAMMA,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_SHORTCONV_CONV,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
+    {LLM_TENSOR_SHORTCONV_INPROJ,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_SHORTCONV_OUTPROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
 };
 
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
@@ -2027,10 +2091,11 @@ bool llm_arch_is_recurrent(const llm_arch & arch) {
 }
 
 bool llm_arch_is_hybrid(const llm_arch & arch) {
-    // List all mamba-attention hybrid models here
     switch (arch) {
         case LLM_ARCH_JAMBA:
         case LLM_ARCH_FALCON_H1:
+        case LLM_ARCH_GRANITE_HYBRID:
+        case LLM_ARCH_LFM2:
             return true;
         default:
             return false;

src/llama-arch.h

@@ -77,6 +77,7 @@ enum llm_arch {
     LLM_ARCH_ARWKV7,
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
+    LLM_ARCH_GRANITE_HYBRID,
     LLM_ARCH_CHAMELEON,
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
@@ -86,6 +87,7 @@ enum llm_arch {
     LLM_ARCH_ERNIE4_5,
     LLM_ARCH_HUNYUAN_MOE,
     LLM_ARCH_SMOLLM3,
+    LLM_ARCH_LFM2,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -158,7 +160,6 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
-    LLM_KV_ATTENTION_LAYER_INDICES,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -227,6 +228,8 @@ enum llm_kv {
 
     LLM_KV_CLASSIFIER_OUTPUT_LABELS,
 
+    LLM_KV_SHORTCONV_L_CACHE,
+
     // deprecated:
     LLM_KV_TOKENIZER_PREFIX_ID,
     LLM_KV_TOKENIZER_SUFFIX_ID,
@@ -396,6 +399,9 @@ enum llm_tensor {
     LLM_TENSOR_POS_NET_ATTN_K,
     LLM_TENSOR_POS_NET_ATTN_V,
     LLM_TENSOR_POS_NET_ATTN_OUT,
+    LLM_TENSOR_SHORTCONV_CONV,
+    LLM_TENSOR_SHORTCONV_INPROJ,
+    LLM_TENSOR_SHORTCONV_OUTPROJ,
 };
 
 enum llm_tensor_layer {

src/llama-chat.cpp

@@ -170,7 +170,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
         // EXAONE-3.0-7.8B-Instruct
         return LLM_CHAT_TEMPLATE_EXAONE_3;
-    } else if (tmpl_contains("rwkv-world")) {
+    } else if (tmpl_contains("rwkv-world") || tmpl_contains("{{- 'User: ' + message['content']|trim + '\\n\\n' -}}")) {
         return LLM_CHAT_TEMPLATE_RWKV_WORLD;
     } else if (tmpl_contains("<|start_of_role|>")) {
         return LLM_CHAT_TEMPLATE_GRANITE;

src/llama-context.cpp

@@ -731,7 +731,8 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     const auto & hparams = model.hparams;
 
-    const int64_t n_embd = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd;
+    const int32_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
     if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, true)) {
@@ -791,10 +792,20 @@ int llama_context::encode(const llama_batch & batch_inp) {
         }
     }
 
+    auto * t_logits = res->get_logits();
     auto * t_embd = res->get_embd_pooled() ? res->get_embd_pooled() : res->get_embd();
 
+    // extract logits
+   if (logits && t_logits) {
+        ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits);
+        GGML_ASSERT(backend_res != nullptr);
+        GGML_ASSERT(logits != nullptr);
+
+        ggml_backend_tensor_get_async(backend_res, t_logits, logits, 0, n_tokens*n_vocab*sizeof(float));
+    }
+
     // extract embeddings
-    if (t_embd) {
+    if (embd && t_embd) {
         ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd);
         GGML_ASSERT(backend_embd != nullptr);
 

src/llama-hparams.cpp

@@ -71,6 +71,11 @@ uint32_t llama_hparams::n_embd_r() const {
         return token_shift_count * n_embd;
     }
 
+    if (n_shortconv_l_cache != 0) {
+        // for LFM2 models
+        return n_embd * (n_shortconv_l_cache - 1);
+    }
+
     // TODO: maybe support other convolution strides than 1
     // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed
     // Corresponds to Mamba's conv_states size

src/llama-hparams.h

@@ -55,6 +55,8 @@ struct llama_hparams {
     struct llama_hparams_posnet   posnet;
     struct llama_hparams_convnext convnext;
 
+    uint32_t n_shortconv_l_cache  = 0;
+
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_arr;
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;

src/llama-model.cpp

@@ -40,17 +40,21 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_190M:          return "190M";
         case LLM_TYPE_220M:          return "220M";
         case LLM_TYPE_250M:          return "250M";
+        case LLM_TYPE_256M:          return "256M";
         case LLM_TYPE_270M:          return "270M";
         case LLM_TYPE_335M:          return "335M";
+        case LLM_TYPE_350M:          return "350M";
         case LLM_TYPE_410M:          return "410M";
         case LLM_TYPE_450M:          return "450M";
         case LLM_TYPE_475M:          return "475M";
+        case LLM_TYPE_700M:          return "700M";
         case LLM_TYPE_770M:          return "770M";
         case LLM_TYPE_780M:          return "780M";
         case LLM_TYPE_0_3B:          return "0.3B";
         case LLM_TYPE_0_5B:          return "0.5B";
         case LLM_TYPE_0_6B:          return "0.6B";
         case LLM_TYPE_1B:            return "1B";
+        case LLM_TYPE_1_2B:          return "1.2B";
         case LLM_TYPE_1_3B:          return "1.3B";
         case LLM_TYPE_1_4B:          return "1.4B";
         case LLM_TYPE_1_5B:          return "1.5B";
@@ -581,6 +585,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                         case 22: type = LLM_TYPE_1B; break;
                         case 26: type = LLM_TYPE_3B; break;
                         case 28: type = LLM_TYPE_3B; break; // Llama 3.2 3B
+                        case 30: type = LLM_TYPE_256M; break; // smoldocling 256M
                         // granite uses a vocab with len 49152
                         case 32: type = n_vocab == 49152 ? LLM_TYPE_3B : (n_vocab < 40000 ? LLM_TYPE_7B : LLM_TYPE_8B); break;
                         case 36: type = LLM_TYPE_8B; break; // granite
@@ -1504,6 +1509,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale);
                 ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale);
 
+                // Granite uses rope_finetuned as a switch for rope, so default to true
+                bool rope_finetuned = true;
+                ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
+                hparams.rope_finetuned = rope_finetuned;
+
                 switch (hparams.n_layer) {
                     case 32: type = LLM_TYPE_3B; break;
                     case 40: type = LLM_TYPE_3B; break;
@@ -1511,6 +1521,40 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
 
+                // For Granite MoE Shared
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
+            } break;
+        case LLM_ARCH_GRANITE_HYBRID:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale, /* required */ false);
+                ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, /* required */ false);
+                ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, /* required */ false);
+                ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, /* required */ false);
+
+                ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+                ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+                ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+                ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+                ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
+
+                // Granite uses rope_finetuned as a switch for rope, so default to true
+                bool rope_finetuned = true;
+                ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
+                hparams.rope_finetuned = rope_finetuned;
+
+                // A layer is recurrent IFF the n_head_kv value is set to 0
+                for (uint32_t i = 0; i < hparams.n_layer; ++i) {
+                    hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0;
+                }
+
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    // TODO: Add llm type label (not sure this is useful)
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+
                 // For Granite MoE Shared
                 ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
             } break;
@@ -1622,6 +1666,20 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_LFM2:
+            {
+                ml.get_key(LLM_KV_SHORTCONV_L_CACHE,           hparams.n_shortconv_l_cache);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                for (uint32_t il = 0; il < hparams.n_layer; ++il) {
+                    hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
+                }
+                switch (hparams.n_embd) {
+                    case 1024: type = LLM_TYPE_350M; break;
+                    case 1536: type = LLM_TYPE_700M; break;
+                    case 2048: type = LLM_TYPE_1_2B; break;
+                    default:   type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -3362,6 +3420,99 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         }
                     }
                 } break;
+            case LLM_ARCH_GRANITE_HYBRID:
+                {
+                    // mamba2 Mixer SSM params
+                    // NOTE: int64_t for tensor dimensions
+                    const int64_t d_conv     = hparams.ssm_d_conv;
+                    const int64_t d_inner    = hparams.ssm_d_inner;
+                    const int64_t d_state    = hparams.ssm_d_state;
+                    const int64_t n_ssm_head = hparams.ssm_dt_rank;
+                    const int64_t n_group    = hparams.ssm_n_group;
+                    const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
+
+                    // only an expansion factor of 2 is supported for now
+                    GGML_ASSERT(2 * n_embd == d_inner);
+
+                    // embeddings
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    {
+                        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed, duplicated to allow offloading
+                        if (output == NULL) {
+                            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        // norm
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        if (hparams.is_recurrent(i)) {
+                            // ssm layers
+                            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
+
+                            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
+                            layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, TENSOR_NOT_REQUIRED);
+
+                            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, 0);
+
+                            // no "weight" suffix for these
+                            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, 0);
+                            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, 0);
+
+                            layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
+
+                            // out_proj
+                            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+                        } else {
+                            // attention layers (with optional bias)
+                            const int64_t n_head_i = hparams.n_head(i);
+                            const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);
+                            const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head_i}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa_i}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa_i}, 0);
+                            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},         TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_k_gqa_i}, TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_v_gqa_i}, TENSOR_NOT_REQUIRED);
+                            layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd},         TENSOR_NOT_REQUIRED);
+                        }
+
+                        // feed forward (w/ optional biases)
+                        if (n_expert > 0) {
+                            // MoE FFN
+                            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+                            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+
+                            // For Granite MoE Shared
+                            if (hparams.n_ff_shexp > 0) {
+                                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
+                            }
+                        } else {
+                            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+                            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+                        }
+                    }
+                } break;
             case LLM_ARCH_XVERSE:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -4772,6 +4923,39 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
                     }
                 } break;
+            case LLM_ARCH_LFM2:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+                    tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+                        // ffn is same for transformer and conv layers
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+                        // for operator_norm
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        if (!hparams.is_recurrent(i)) {
+                            layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+                            layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+                            GGML_ASSERT(n_embd_v_gqa == n_embd_k_gqa);
+
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, hparams.n_embd_k_gqa(i)}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, hparams.n_embd_v_gqa(i)}, 0);
+
+                            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+                        } else {
+                            layer.shortconv.conv     = create_tensor(tn(LLM_TENSOR_SHORTCONV_CONV,    "weight", i), {hparams.n_shortconv_l_cache, n_embd}, 0);
+                            layer.shortconv.in_proj  = create_tensor(tn(LLM_TENSOR_SHORTCONV_INPROJ,  "weight", i), {n_embd, 3 * n_embd}, 0);
+                            layer.shortconv.out_proj = create_tensor(tn(LLM_TENSOR_SHORTCONV_OUTPROJ, "weight", i), {n_embd, n_embd}, 0);
+                        }
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -5024,7 +5208,8 @@ void llama_model::print_info() const {
     if (arch == LLM_ARCH_MAMBA ||
         arch == LLM_ARCH_MAMBA2 ||
         arch == LLM_ARCH_JAMBA ||
-        arch == LLM_ARCH_FALCON_H1) {
+        arch == LLM_ARCH_FALCON_H1 ||
+        arch == LLM_ARCH_GRANITE_HYBRID) {
         LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
         LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
         LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
@@ -5079,7 +5264,8 @@ void llama_model::print_info() const {
 
     if (arch == LLM_ARCH_MINICPM ||
         arch == LLM_ARCH_GRANITE ||
-        arch == LLM_ARCH_GRANITE_MOE) {
+        arch == LLM_ARCH_GRANITE_MOE ||
+        arch == LLM_ARCH_GRANITE_HYBRID) {
         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
@@ -13795,13 +13981,11 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base {
     }
 };
 
-
 struct llm_build_granite : public llm_graph_context {
     llm_build_granite(
         const llama_model & model,
         const llm_graph_params & params,
-        ggml_cgraph * gf,
-        const bool use_rope = true)
+        ggml_cgraph * gf)
         : llm_graph_context(params) {
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -13816,14 +14000,12 @@ struct llm_build_granite : public llm_graph_context {
 
         // inp_pos - built only if rope enabled
         ggml_tensor * inp_pos = nullptr;
-        if (use_rope) {
+        if (hparams.rope_finetuned) {
             inp_pos = build_inp_pos();
         }
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
         for (int il = 0; il < n_layer; ++il) {
@@ -13836,128 +14018,17 @@ struct llm_build_granite : public llm_graph_context {
             cb(cur, "attn_norm", il);
 
             // self-attention
-            {
-                // compute Q and K and (optionally) RoPE them
-                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
-                if (use_rope) {
-                    ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-                    Qcur = ggml_rope_ext(
-                            ctx0, Qcur, inp_pos, rope_factors,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-
-                    Kcur = ggml_rope_ext(
-                            ctx0, Kcur, inp_pos, rope_factors,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                cur = build_attn(inp_attn, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
-                cb(cur, "attn_out", il);
-            }
+            cur = build_attention_layer(
+                gf, cur, inp_pos, inp_attn,
+                model, n_embd_head, il);
 
             if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
 
-            // For Granite architectures - scale residual
-            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network (non-MoE)
-            if (model.layers[il].ffn_gate_inp == nullptr) {
-
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = build_ffn(cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                cb(cur, "ffn_out", il);
-
-            } else {
-                // MoE branch
-                cur = build_norm(ffn_inp,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(cur, "ffn_norm", il);
-
-                ggml_tensor * moe_out = build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, true,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        il);
-                cb(moe_out, "ffn_moe_out", il);
-
-                // For Granite MoE Shared
-                if (hparams.n_ff_shexp > 0) {
-                    ggml_tensor * ffn_shexp = build_ffn(cur,
-                        model.layers[il].ffn_up_shexp,   NULL, NULL,
-                        model.layers[il].ffn_gate_shexp, NULL, NULL,
-                        model.layers[il].ffn_down_shexp, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, il);
-                    cb(ffn_shexp, "ffn_shexp", il);
-
-                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                    cb(cur, "ffn_out", il);
-                } else {
-                    cur = moe_out;
-                }
-            }
-
-            // For Granite architectures - scale residual
-            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = build_cvec(cur, il);
-            cb(cur, "l_out", il);
+            // ffn
+            cur = build_layer_ffn(cur, inpSA, model, il);
 
             // input for next layer
             inpL = cur;
@@ -13982,6 +14053,370 @@ struct llm_build_granite : public llm_graph_context {
 
         ggml_build_forward_expand(gf, cur);
     }
+
+    ggml_tensor * build_attention_layer(
+              ggml_cgraph                     * gf,
+              ggml_tensor                     * cur,
+              ggml_tensor                     * inp_pos,
+              llm_graph_input_attn_kv_unified * inp_attn,
+        const llama_model                     & model,
+        const int64_t                           n_embd_head,
+        const int                               il) {
+
+        // compute Q and K and (optionally) RoPE them
+        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+        cb(Qcur, "Qcur", il);
+        if (model.layers[il].bq) {
+            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur", il);
+        }
+
+        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+        cb(Kcur, "Kcur", il);
+        if (model.layers[il].bk) {
+            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur", il);
+        }
+
+        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+        cb(Vcur, "Vcur", il);
+        if (model.layers[il].bv) {
+            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur", il);
+        }
+
+        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il),    n_tokens);
+        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+        const bool use_rope = hparams.rope_finetuned;
+        if (use_rope) {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+        }
+
+        cb(Qcur, "Qcur", il);
+        cb(Kcur, "Kcur", il);
+        cb(Vcur, "Vcur", il);
+
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+        cur = build_attn(inp_attn, gf,
+                model.layers[il].wo, model.layers[il].bo,
+                Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+        return cur;
+    }
+
+    ggml_tensor * build_layer_ffn(
+              ggml_tensor       * cur,
+              ggml_tensor       * inpSA,
+        const llama_model       & model,
+        const int                 il) {
+
+        // For Granite architectures - scale residual
+        if (hparams.f_residual_scale) {
+            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network (non-MoE)
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+                    cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+                    cb(cur, "ffn_out", il);
+
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+                    cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // For Granite MoE Shared
+            if (hparams.n_ff_shexp > 0) {
+                ggml_tensor * ffn_shexp = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            } else {
+                cur = moe_out;
+            }
+        }
+
+        // For Granite architectures - scale residual
+        if (hparams.f_residual_scale) {
+            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        return cur;
+    }
+};
+
+struct llm_build_granite_hybrid : public llm_graph_context_mamba {
+
+    llm_build_granite_hybrid(
+                 const llama_model & model,
+            const llm_graph_params & params,
+                       ggml_cgraph * gf) :
+        llm_graph_context_mamba(params) {
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        auto * inp = build_inp_mem_hybrid();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        // Positional embeddings populated if rope enabled
+        ggml_tensor * inp_pos = nullptr;
+        if (hparams.rope_finetuned) {
+            inp_pos = build_inp_pos();
+        }
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            if (hparams.is_recurrent(il)) {
+                // ssm layer //
+                cur = build_mamba2_layer(inp->get_recr(), gf, cur, model, ubatch, il);
+            } else {
+                // attention layer //
+                cur = build_attention_layer(
+                    gf, cur, inp_pos, inp->get_attn(), model,
+                    n_embd_head, il);
+            }
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // ffn
+            cur = build_layer_ffn(cur, inpSA, model, il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        // For Granite architectures - scale logits
+        if (hparams.f_logit_scale) {
+            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+        }
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+
+    ggml_tensor * build_attention_layer(
+              ggml_cgraph                     * gf,
+              ggml_tensor                     * cur,
+              ggml_tensor                     * inp_pos,
+              llm_graph_input_attn_kv_unified * inp_attn,
+        const llama_model                     & model,
+        const int64_t                           n_embd_head,
+        const int                               il) {
+
+        // compute Q and K and (optionally) RoPE them
+        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+        cb(Qcur, "Qcur", il);
+        if (model.layers[il].bq) {
+            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur", il);
+        }
+
+        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+        cb(Kcur, "Kcur", il);
+        if (model.layers[il].bk) {
+            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur", il);
+        }
+
+        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+        cb(Vcur, "Vcur", il);
+        if (model.layers[il].bv) {
+            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur", il);
+        }
+
+        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il),    n_tokens);
+        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+        const bool use_rope = hparams.rope_finetuned;
+        if (use_rope) {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+        }
+
+        cb(Qcur, "Qcur", il);
+        cb(Kcur, "Kcur", il);
+        cb(Vcur, "Vcur", il);
+
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+        cur = build_attn(inp_attn, gf,
+                model.layers[il].wo, model.layers[il].bo,
+                Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+        return cur;
+    }
+
+    ggml_tensor * build_layer_ffn(
+              ggml_tensor       * cur,
+              ggml_tensor       * inpSA,
+        const llama_model       & model,
+        const int                 il) {
+
+        // For Granite architectures - scale residual
+        if (hparams.f_residual_scale) {
+            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network (non-MoE)
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+                    cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+                    cb(cur, "ffn_out", il);
+
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+                    cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // For Granite MoE Shared
+            if (hparams.n_ff_shexp > 0) {
+                ggml_tensor * ffn_shexp = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            } else {
+                cur = moe_out;
+            }
+        }
+
+        // For Granite architectures - scale residual
+        if (hparams.f_residual_scale) {
+            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        return cur;
+    }
 };
 
 // ref: https://github.com/facebookresearch/chameleon
@@ -15474,6 +15909,163 @@ struct llm_build_smollm3 : public llm_graph_context {
     }
 };
 
+struct llm_build_lfm2 : public llm_graph_context {
+    const llama_model & model;
+
+    llm_build_lfm2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params), model(model) {
+
+        ggml_tensor * cur = build_inp_embd(model.tok_embd);
+        cb(cur, "model.embed_tokens", -1);
+
+        ggml_tensor * inp_pos     = build_inp_pos();
+        auto        * inp_hybrid  = build_inp_mem_hybrid();
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            auto * prev_cur = cur;
+            cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "model.layers.{}.operator_norm", il);
+
+            cur = hparams.is_recurrent(il) ?
+                build_shortconv_block(gf, cur, inp_hybrid->get_recr(), il) :
+                build_attn_block(gf, cur, inp_pos, inp_hybrid->get_attn(), il) ;
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur      = ggml_get_rows(ctx0,      cur, inp_out_ids);
+                prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids);
+            }
+
+            cur = ggml_add(ctx0, prev_cur, cur);
+            cur = ggml_add(ctx0, cur, build_feed_forward(cur, il));
+        }
+
+        cur = build_norm(cur, model.tok_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "model.embedding_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head is tied with embeddings
+        cur = build_lora_mm(model.tok_embd, cur);
+        cb(cur, "lm_head", -1);
+
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+
+    ggml_tensor * build_feed_forward(ggml_tensor * cur,
+                                     int           il) const {
+        cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "model.layers.{}.ffn_norm", il);
+
+        GGML_ASSERT(!model.layers[il].ffn_up_b);
+        GGML_ASSERT(!model.layers[il].ffn_gate_b);
+        GGML_ASSERT(!model.layers[il].ffn_down_b);
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "model.layers.{}.feed_forward.w2", il);
+
+        return cur;
+    }
+
+    ggml_tensor * build_attn_block(ggml_cgraph                     * gf,
+                                   ggml_tensor                     * cur,
+                                   ggml_tensor                     * inp_pos,
+                                   llm_graph_input_attn_kv_unified * inp_attn,
+                                   int                               il) const {
+        GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
+        auto const n_embd_head = hparams.n_embd_head_v;
+        auto const n_head_kv = hparams.n_head_kv(il);
+
+        auto * q = build_lora_mm(model.layers[il].wq, cur);
+        cb(q, "model.layers.{}.self_attn.q_proj", il);
+        auto * k = build_lora_mm(model.layers[il].wk, cur);
+        cb(k, "model.layers.{}.self_attn.k_proj", il);
+        auto * v = build_lora_mm(model.layers[il].wv, cur);
+        cb(v, "model.layers.{}.self_attn.v_proj", il);
+
+        q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head,    n_tokens);
+        k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens);
+        v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens);
+
+        // qk norm
+        q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+        cb(q, "model.layers.{}.self_attn.q_layernorm", il);
+        k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+        cb(k, "model.layers.{}.self_attn.k_layernorm", il);
+
+        // RoPE
+        q = ggml_rope_ext(
+                ctx0, q, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+        k = ggml_rope_ext(
+                ctx0, k, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+
+        cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL,
+                q, k, v, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+
+        cb(cur, "model.layers.{}.self_attn.out_proj", il);
+
+        return cur;
+    }
+
+    ggml_tensor * build_shortconv_block(ggml_cgraph        * gf,
+                                        ggml_tensor        * cur,
+                                        llm_graph_input_rs * inp_recr,
+                                        int                il) {
+        const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx)->get_recr();
+
+        auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);
+        cb(bcx, "model.layers.{}.conv.in_proj", il);
+
+        constexpr auto n_chunks = 3;
+        GGML_ASSERT(bcx->ne[0] % n_chunks == 0);
+        auto const chunk_size = bcx->ne[0] / n_chunks;
+        auto * b = ggml_view_2d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->nb[1], 0 * chunk_size * ggml_element_size(bcx));
+        auto * c = ggml_view_2d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->nb[1], 1 * chunk_size * ggml_element_size(bcx));
+        auto * x = ggml_view_2d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->nb[1], 2 * chunk_size * ggml_element_size(bcx));
+
+        auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));
+
+        // read conv state directly, with build_rs generation is slower
+        ggml_tensor * conv_state = mctx_cur->get_r_l(il);
+        const int64_t n_seqs  = ubatch.n_seqs;
+        ggml_tensor * conv = build_rs(inp_recr, gf, conv_state, hparams.n_embd_r(), n_seqs);
+        conv = ggml_reshape_3d(ctx0, conv_state, hparams.n_shortconv_l_cache - 1, hparams.n_embd, n_seqs);
+
+        bx = ggml_concat(ctx0, conv, bx, 0);
+        GGML_ASSERT(bx->ne[0] > conv->ne[0]);
+
+        auto * new_conv = ggml_view_2d(ctx0, bx, conv->ne[0], bx->ne[1], bx->nb[1], (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));
+        GGML_ASSERT(ggml_are_same_shape(conv, new_conv));
+
+        // write conv state
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv, conv_state));
+
+        auto * conv_kernel = model.layers[il].shortconv.conv;
+        GGML_ASSERT(hparams.n_shortconv_l_cache > 0);
+
+        // construct ssm_conv op
+        ggml_tensor * conv_out = ggml_ssm_conv(ctx0, bx, conv_kernel);
+        cb(conv_out, "model.layers.{}.conv.conv", il);
+
+        auto * y = ggml_mul(ctx0, c, conv_out);
+
+        y = build_lora_mm(model.layers[il].shortconv.out_proj, y);
+        cb(y, "model.layers.{}.conv.out_proj", il);
+
+        return y;
+    }
+};
+
 llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
     llama_memory_i * res;
 
@@ -15832,6 +16424,10 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_granite>(*this, params, gf);
             } break;
+        case LLM_ARCH_GRANITE_HYBRID:
+            {
+                llm = std::make_unique<llm_build_granite_hybrid>(*this, params, gf);
+            } break;
         case LLM_ARCH_CHAMELEON:
             {
                 llm = std::make_unique<llm_build_chameleon>(*this, params, gf);
@@ -15872,6 +16468,10 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_falcon_h1>(*this, params, gf);
             } break;
+        case LLM_ARCH_LFM2:
+            {
+                llm = std::make_unique<llm_build_lfm2>(*this, params, gf);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -16021,6 +16621,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_GLM4:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
+        case LLM_ARCH_GRANITE_HYBRID:
         case LLM_ARCH_CHAMELEON:
         case LLM_ARCH_BAILINGMOE:
         case LLM_ARCH_NEO_BERT:
@@ -16064,6 +16665,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_MINICPM3:
         case LLM_ARCH_DOTS1:
         case LLM_ARCH_HUNYUAN_MOE:
+        case LLM_ARCH_LFM2:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:

src/llama-model.h

@@ -32,17 +32,21 @@ enum llm_type {
     LLM_TYPE_190M,
     LLM_TYPE_220M,
     LLM_TYPE_250M,
+    LLM_TYPE_256M,
     LLM_TYPE_270M,
     LLM_TYPE_335M,
+    LLM_TYPE_350M,
     LLM_TYPE_410M,
     LLM_TYPE_450M,
     LLM_TYPE_475M,
+    LLM_TYPE_700M,
     LLM_TYPE_770M,
     LLM_TYPE_780M,
     LLM_TYPE_0_3B,
     LLM_TYPE_0_5B,
     LLM_TYPE_0_6B,
     LLM_TYPE_1B,
+    LLM_TYPE_1_2B,
     LLM_TYPE_1_3B,
     LLM_TYPE_1_4B,
     LLM_TYPE_1_5B,
@@ -154,6 +158,12 @@ struct llama_layer_convnext {
     struct ggml_tensor * gamma = nullptr;
 };
 
+struct llama_layer_shortconv {
+    struct ggml_tensor * in_proj  = nullptr;
+    struct ggml_tensor * conv     = nullptr;
+    struct ggml_tensor * out_proj = nullptr;
+};
+
 struct llama_layer {
     // normalization
     struct ggml_tensor * attn_norm       = nullptr;
@@ -340,6 +350,8 @@ struct llama_layer {
     struct llama_layer_posnet posnet;
 
     struct llama_layer_convnext convnext;
+
+    struct llama_layer_shortconv shortconv;
 };
 
 struct llama_model {

src/llama-quant.cpp

@@ -844,6 +844,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         // do not quantize Mamba's small yet 2D weights
         // NOTE: can't use LLM_TN here because the layer number is not known
         quantize &= name.find("ssm_conv1d.weight") == std::string::npos;
+        quantize &= name.find("shortconv.conv.weight") == std::string::npos;
 
         // do not quantize RWKV's small yet 2D weights
         quantize &= name.find("time_mix_first.weight") == std::string::npos;
@@ -883,8 +884,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                         if (std::regex pattern(tname); std::regex_search(tensor_name, pattern)) {
                             if  (qtype != new_type) {
                                 LLAMA_LOG_DEBUG("(overriding %s) ", ggml_type_name(new_type));
-                                new_type = qtype;
-                                break; // if two or more types are specified for the tensor, first match wins
+                                new_type = qtype; // if two or more types are specified for the same tensor, the last match wins
                             }
                         }
                     }

src/llama-vocab.cpp

@@ -1524,7 +1524,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "llama-bpe"||
                     tokenizer_pre == "falcon3"  ||
                     tokenizer_pre == "falcon-h1" ||
-                    tokenizer_pre == "pixtral") {
+                    tokenizer_pre == "pixtral"  ||
+                    tokenizer_pre == "midm-2.0" ||
+                    tokenizer_pre == "lfm2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
                 ignore_merges = true;
                 add_bos = true;
@@ -1846,6 +1848,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<EOT>"
                         || t.first == "_<EOT>"
                         || t.first == "<｜end▁of▁sentence｜>" // DeepSeek
+                        || t.first == "<end_of_utterance>" // smoldocling
                    ) {
                     special_eot_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2005,6 +2008,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<EOT>"
                     || t.first == "_<EOT>"
                     || t.first == "<|end_of_text|>"
+                    || t.first == "<end_of_utterance>" // smoldocling
                ) {
                 special_eog_ids.insert(t.second);
                 if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {

src/llama-vocab.h

@@ -6,6 +6,47 @@
 #include <vector>
 #include <memory>
 
+// pre-tokenization types
+enum llama_vocab_pre_type {
+    LLAMA_VOCAB_PRE_TYPE_DEFAULT        = 0,
+    LLAMA_VOCAB_PRE_TYPE_LLAMA3         = 1,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM   = 2,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,
+    LLAMA_VOCAB_PRE_TYPE_FALCON         = 4,
+    LLAMA_VOCAB_PRE_TYPE_MPT            = 5,
+    LLAMA_VOCAB_PRE_TYPE_STARCODER      = 6,
+    LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
+    LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
+    LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
+    LLAMA_VOCAB_PRE_TYPE_STABLELM2      = 10,
+    LLAMA_VOCAB_PRE_TYPE_QWEN2          = 11,
+    LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,
+    LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
+    LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,
+    LLAMA_VOCAB_PRE_TYPE_PORO           = 15,
+    LLAMA_VOCAB_PRE_TYPE_CHATGLM3       = 16,
+    LLAMA_VOCAB_PRE_TYPE_CHATGLM4       = 17,
+    LLAMA_VOCAB_PRE_TYPE_VIKING         = 18,
+    LLAMA_VOCAB_PRE_TYPE_JAIS           = 19,
+    LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
+    LLAMA_VOCAB_PRE_TYPE_SMOLLM         = 21,
+    LLAMA_VOCAB_PRE_TYPE_CODESHELL      = 22,
+    LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
+    LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
+    LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
+    LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
+    LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 28,
+    LLAMA_VOCAB_PRE_TYPE_GPT4O          = 29,
+    LLAMA_VOCAB_PRE_TYPE_SUPERBPE       = 30,
+    LLAMA_VOCAB_PRE_TYPE_TRILLION       = 31,
+    LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
+    LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
+    LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
+    LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
+    LLAMA_VOCAB_PRE_TYPE_HUNYUAN        = 36,
+};
+
 struct LLM_KV;
 struct llama_model_loader;
 

tests/test-backend-ops.cpp

@@ -317,10 +317,11 @@ enum test_mode {
     MODE_TEST,
     MODE_PERF,
     MODE_GRAD,
+    MODE_SUPPORT,
 };
 
 // Output format support similar to llama-bench
-enum output_formats { CONSOLE, SQL };
+enum output_formats { CONSOLE, SQL, CSV };
 
 static const char * output_format_str(output_formats format) {
     switch (format) {
@@ -328,6 +329,8 @@ static const char * output_format_str(output_formats format) {
             return "console";
         case SQL:
             return "sql";
+        case CSV:
+            return "csv";
         default:
             GGML_ABORT("invalid output format");
     }
@@ -338,6 +341,8 @@ static bool output_format_from_str(const std::string & s, output_formats & forma
         format = CONSOLE;
     } else if (s == "sql") {
         format = SQL;
+    } else if (s == "csv") {
+        format = CSV;
     } else {
         return false;
     }
@@ -360,6 +365,8 @@ struct test_result {
     double      bandwidth_gb_s;
     size_t      memory_kb;
     int         n_runs;
+    std::string device_description;
+    std::string backend_reg_name;
 
     test_result() {
         // Initialize with default values
@@ -384,7 +391,7 @@ struct test_result {
     test_result(const std::string & backend_name, const std::string & op_name, const std::string & op_params,
                 const std::string & test_mode, bool supported, bool passed, const std::string & error_message = "",
                 double time_us = 0.0, double flops = 0.0, double bandwidth_gb_s = 0.0, size_t memory_kb = 0,
-                int n_runs = 0) :
+                int n_runs = 0, const std::string & device_description = "", const std::string & backend_reg_name = "") :
         backend_name(backend_name),
         op_name(op_name),
         op_params(op_params),
@@ -396,7 +403,9 @@ struct test_result {
         flops(flops),
         bandwidth_gb_s(bandwidth_gb_s),
         memory_kb(memory_kb),
-        n_runs(n_runs) {
+        n_runs(n_runs),
+        device_description(device_description),
+        backend_reg_name(backend_reg_name) {
         // Set test time
         time_t t = time(NULL);
         char   buf[32];
@@ -410,7 +419,8 @@ struct test_result {
     static const std::vector<std::string> & get_fields() {
         static const std::vector<std::string> fields = {
             "test_time", "build_commit",  "backend_name", "op_name", "op_params",      "test_mode", "supported",
-            "passed",    "error_message", "time_us",      "flops",   "bandwidth_gb_s", "memory_kb", "n_runs"
+            "passed",    "error_message", "time_us",      "flops",   "bandwidth_gb_s", "memory_kb", "n_runs",
+            "device_description", "backend_reg_name"
         };
         return fields;
     }
@@ -444,7 +454,9 @@ struct test_result {
                  std::to_string(flops),
                  std::to_string(bandwidth_gb_s),
                  std::to_string(memory_kb),
-                 std::to_string(n_runs) };
+                 std::to_string(n_runs),
+                 device_description,
+                 backend_reg_name };
     }
 };
 
@@ -633,6 +645,8 @@ struct console_printer : public printer {
             print_test_console(result);
         } else if (result.test_mode == "perf") {
             print_perf_console(result);
+        } else if (result.test_mode == "support") {
+            print_support_console(result);
         }
     }
 
@@ -799,6 +813,17 @@ struct console_printer : public printer {
         }
         printf("\n");
     }
+
+    void print_support_console(const test_result & result) {
+        printf("  %s(%s): ", result.op_name.c_str(), result.op_params.c_str());
+        fflush(stdout);
+
+        if (result.supported) {
+            printf("\033[1;32mSUPPORTED\033[0m\n");
+        } else {
+            printf("\033[1;31mNOT SUPPORTED\033[0m\n");
+        }
+    }
 };
 
 struct sql_printer : public printer {
@@ -841,12 +866,39 @@ struct sql_printer : public printer {
     }
 };
 
+struct csv_printer : public printer {
+    void print_header() override {
+        std::vector<std::string> fields = test_result::get_fields();
+        for (size_t i = 0; i < fields.size(); i++) {
+            printf("\"%s\"%s", fields[i].c_str(), i < fields.size() - 1 ? "," : "");
+        }
+        printf("\n");
+    }
+
+    void print_test_result(const test_result & result) override {
+        std::vector<std::string> values = result.get_values();
+        for (size_t i = 0; i < values.size(); i++) {
+            // Escape quotes and wrap in quotes for CSV
+            std::string escaped_value = values[i];
+            size_t pos = 0;
+            while ((pos = escaped_value.find("\"", pos)) != std::string::npos) {
+                escaped_value.replace(pos, 1, "\"\"");
+                pos += 2;
+            }
+            printf("\"%s\"%s", escaped_value.c_str(), i < values.size() - 1 ? "," : "");
+        }
+        printf("\n");
+    }
+};
+
 static std::unique_ptr<printer> create_printer(output_formats format) {
     switch (format) {
         case CONSOLE:
             return std::make_unique<console_printer>();
         case SQL:
             return std::make_unique<sql_printer>();
+        case CSV:
+            return std::make_unique<csv_printer>();
     }
     GGML_ABORT("invalid output format");
 }
@@ -928,7 +980,7 @@ struct test_case {
     std::vector<ggml_tensor *> sentinels;
 
     void add_sentinel(ggml_context * ctx) {
-        if (mode == MODE_PERF || mode == MODE_GRAD) {
+        if (mode == MODE_PERF || mode == MODE_GRAD || mode == MODE_SUPPORT) {
             return;
         }
         ggml_tensor * sentinel = ::ggml_new_tensor_1d(ctx, GGML_TYPE_F32, sentinel_size);
@@ -1153,15 +1205,12 @@ struct test_case {
             return true;
         }
 
-        // check if backends support op
         if (!ggml_backend_supports_op(backend, out)) {
             // Create test result for unsupported performance test
             test_result result(ggml_backend_name(backend), current_op_name, vars(), "perf", false, false,
                                "not supported");
 
-            if (output_printer) {
-                output_printer->print_test_result(result);
-            }
+            output_printer->print_test_result(result);
 
             return true;
         }
@@ -1266,6 +1315,38 @@ struct test_case {
         return true;
     }
 
+    bool eval_support(ggml_backend_t backend, const char * op_name, printer * output_printer) {
+        mode = MODE_SUPPORT;
+
+        static const size_t graph_nodes = 8192;
+
+        ggml_init_params params = {
+            /* .mem_size = */ ggml_tensor_overhead()*128 + ggml_graph_overhead_custom(graph_nodes, false),
+            /* .mem_base = */ NULL,
+            /* .no_alloc = */ true,
+        };
+        ggml_context_ptr ctx(ggml_init(params)); // smart ptr
+        GGML_ASSERT(ctx);
+
+        ggml_tensor * out             = build_graph(ctx.get());
+        std::string   current_op_name = op_desc(out);
+        if (op_name != nullptr && current_op_name != op_name) {
+            return true;
+        }
+
+        bool supported = ggml_backend_supports_op(backend, out);
+
+        std::string device_desc = ggml_backend_dev_description(ggml_backend_get_device(backend));
+        std::string backend_reg_name = ggml_backend_reg_name(ggml_backend_dev_backend_reg(ggml_backend_get_device(backend)));
+
+        test_result result(ggml_backend_name(backend), current_op_name, vars(), "support", supported, supported,
+                           supported ? "yes" : "no", 0.0, 0.0, 0.0, 0, 0, device_desc, backend_reg_name);
+
+        output_printer->print_test_result(result);
+
+        return true;
+    }
+
     bool eval_grad(ggml_backend_t backend, const char * op_name, printer * output_printer) {
         mode = MODE_GRAD;
         const std::vector<float> expect = grad_expect();
@@ -4033,6 +4114,32 @@ struct test_pad_reflect_1d : public test_case {
     }
 };
 
+// GGML_OP_ROLL
+struct test_roll : public test_case {
+    const int shift0;
+    const int shift1;
+    const int shift3;
+    const int shift4;
+
+    std::string vars() override {
+        return VARS_TO_STR4(shift0, shift1, shift3, shift4);
+    }
+
+    test_roll(int shift0 = 3, int shift1 = -2, int shift3 = 1, int shift4 = -1)
+        : shift0(shift0), shift1(shift1), shift3(shift3), shift4(shift4) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        int64_t ne[4] = {10, 5, 4, 3};
+        ggml_tensor * a = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+        ggml_set_name(a, "a");
+
+        ggml_tensor * out = ggml_roll(ctx, a, shift0, shift1, shift3, shift4);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+};
+
 // GGML_OP_ARANGE
 struct test_arange : public test_case {
     const ggml_type type;
@@ -5063,9 +5170,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, {64, 5, 4, 3}, 1e-12f));
 
-    test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 1, 1}, {4, 1536, 1, 1}));
-    test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {8, 1536, 1, 1}, {4, 1536, 1, 1}));
-    test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 4, 1}, {4, 1536, 1, 1}));
+    for (int64_t d_conv : {3, 4}) {
+        for (int64_t d_inner: {1024, 1536, 2048}) {
+            test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, d_inner, 1, 1}, {d_conv, d_inner, 1, 1}));
+            test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {8, d_inner, 1, 1}, {d_conv, d_inner, 1, 1}));
+            test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, d_inner, 4, 1}, {d_conv, d_inner, 1, 1}));
+        }
+    }
 
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 16, 1, 1024, 1, 32, 4)); // Mamba-1
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 16, 2, 32, 4)); // Mamba-2
@@ -5403,6 +5514,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_acc());
     test_cases.emplace_back(new test_pad());
     test_cases.emplace_back(new test_pad_reflect_1d());
+    test_cases.emplace_back(new test_roll());
     test_cases.emplace_back(new test_arange());
     test_cases.emplace_back(new test_timestep_embedding());
     test_cases.emplace_back(new test_leaky_relu());
@@ -5599,17 +5711,27 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         return true;
     }
 
+    if (mode == MODE_SUPPORT) {
+        auto test_cases = make_test_cases_eval();
+        filter_test_cases(test_cases, params_filter);
+        for (auto & test : test_cases) {
+            test->eval_support(backend, op_name, output_printer);
+        }
+        return true;
+    }
+
     GGML_ABORT("fatal error");
 }
 
 static void usage(char ** argv) {
-    printf("Usage: %s [mode] [-o <op>] [-b <backend>] [-p <params regex>] [--output <console|sql>]\n", argv[0]);
+    printf("Usage: %s [mode] [-o <op>] [-b <backend>] [-p <params regex>] [--output <console|sql|csv>]\n", argv[0]);
     printf("    valid modes:\n");
     printf("      - test (default, compare with CPU backend for correctness)\n");
     printf("      - grad (compare gradients from backpropagation with method of finite differences)\n");
     printf("      - perf (performance evaluation)\n");
+    printf("      - support (probe backend operation support)\n");
     printf("    op names for -o are as given by ggml_op_desc() (e.g. ADD, MUL_MAT, etc)\n");
-    printf("    --output specifies output format (default: console)\n");
+    printf("    --output specifies output format (default: console, options: console, sql, csv)\n");
 }
 
 int main(int argc, char ** argv) {
@@ -5626,6 +5748,8 @@ int main(int argc, char ** argv) {
             mode = MODE_PERF;
         } else if (strcmp(argv[i], "grad") == 0) {
             mode = MODE_GRAD;
+        } else if (strcmp(argv[i], "support") == 0) {
+            mode = MODE_SUPPORT;
         } else if (strcmp(argv[i], "-o") == 0) {
             if (i + 1 < argc) {
                 op_name_filter = argv[++i];

tools/run/CMakeLists.txt

@@ -7,8 +7,7 @@ if (LLAMA_CURL)
     find_package(CURL REQUIRED)
     target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
     include_directories(${CURL_INCLUDE_DIRS})
-    find_library(CURL_LIBRARY curl REQUIRED)
-    set(LLAMA_RUN_EXTRA_LIBS ${LLAMA_RUN_EXTRA_LIBS} ${CURL_LIBRARY})
+    set(LLAMA_RUN_EXTRA_LIBS ${LLAMA_RUN_EXTRA_LIBS} ${CURL_LIBRARIES})
 endif ()
 
 install(TARGETS ${TARGET} RUNTIME)

tools/server/server.cpp

@@ -2581,12 +2581,14 @@ struct server_context {
                 continue;
             }
 
-            const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-            if (embd == NULL) {
+            const float * embd = nullptr;
+            if (llama_pooling_type(slot.ctx) == LLAMA_POOLING_TYPE_NONE) {
                 embd = llama_get_embeddings_ith(ctx, i);
+            } else {
+                embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
             }
 
-            if (embd == NULL) {
+            if (embd == nullptr) {
                 SLT_ERR(slot, "failed to get embeddings, token = %d, seq_id = %d\n", batch.token[i], batch.seq_id[i][0]);
 
                 res->embedding.push_back(std::vector<float>(n_embd, 0.0f));
@@ -2594,12 +2596,12 @@ struct server_context {
             }
 
             // normalize only when there is pooling
-            // TODO: configurable
             if (llama_pooling_type(slot.ctx) != LLAMA_POOLING_TYPE_NONE) {
                 common_embd_normalize(embd, embd_res.data(), n_embd, 2);
                 res->embedding.push_back(embd_res);
+                break;
             } else {
-                res->embedding.push_back({ embd, embd + n_embd });
+                res->embedding.emplace_back(embd, embd + n_embd);
             }
         }
 

tools/server/utils.hpp

@@ -11,6 +11,8 @@
 
 // increase max payload length to allow use of larger context size
 #define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
+// increase backlog size to avoid connection resets for >> 1 slots
+#define CPPHTTPLIB_LISTEN_BACKLOG 512
 // disable Nagle's algorithm
 #define CPPHTTPLIB_TCP_NODELAY true
 #include <cpp-httplib/httplib.h>