Merge branch 'ggml-org:master' into mradermacher

Author: nicoboss

.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."

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

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)
@@ -112,13 +111,13 @@ 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
         CONFIGURE_COMMAND ""
-        BUILD_COMMAND cargo build --release
+        BUILD_COMMAND cargo build --release --package llguidance
         INSTALL_COMMAND ""
         BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME} ${LLGUIDANCE_PATH}/llguidance.h
         UPDATE_COMMAND ""

convert_hf_to_gguf.py

@@ -300,6 +300,7 @@ def prepare_tensors(self):
                             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 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         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")
@@ -4890,6 +4897,9 @@ def __init__(self, dir_model: Path, *args, **kwargs):
             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"]
@@ -4912,32 +4922,29 @@ def set_vocab(self):
             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
 
         # Fail early for models which don't have a block expansion factor of 2
         # 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 +4969,7 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
             # (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 +6456,148 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
                 (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, 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_GATE_SHEXP, bid), gate),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+                (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 +6821,7 @@ def __init__(self, *args, **kwargs):
         # 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 +7077,50 @@ def set_vocab(self):
             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 @@ class TOKENIZER_TYPE(IntEnum):
     {"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