llama : make llm_tokenizer more private

ggml-ci

CMakeLists.txt

@@ -82,11 +82,11 @@ set(GGML_FATAL_WARNINGS     ${LLAMA_FATAL_WARNINGS})
 
 # change the default for these ggml options
 if (NOT DEFINED GGML_LLAMAFILE)
-    set(GGML_LLAMAFILE ON)
+    set(GGML_LLAMAFILE_DEFAULT ON)
 endif()
 
-if (NOT DEFINED GGML_CUDA_USE_GRAPHS)
-    set(GGML_CUDA_USE_GRAPHS ON)
+if (NOT DEFINED GGML_CUDA_GRAPHS)
+    set(GGML_CUDA_GRAPHS_DEFAULT ON)
 endif()
 
 # transition helpers

Makefile

@@ -619,7 +619,7 @@ ifdef GGML_CUDA
 			CUDA_PATH ?= /usr/local/cuda
 		endif
 
-		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
 		MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
 		MK_NVCCFLAGS += -use_fast_math
 	endif # GGML_MUSA

README.md

@@ -77,6 +77,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [OLMoE](https://huggingface.co/allenai/OLMoE-1B-7B-0924)
 - [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
 - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
 - [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)

common/arg.cpp

@@ -1312,7 +1312,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
         [](gpt_params & params, int value) {
             params.n_parallel = value;
         }
-    ));
+    ).set_env("LLAMA_ARG_N_PARALLEL"));
     add_opt(llama_arg(
         {"-ns", "--sequences"}, "N",
         format("number of sequences to decode (default: %d)", params.n_sequences),

convert_hf_to_gguf.py

@@ -132,12 +132,14 @@ class Model:
     def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
         tensor_names_from_parts: set[str] = set()
 
-        if len(self.part_names) > 1:
+        index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
+        index_name += ".index.json"
+        index_file = self.dir_model / index_name
+
+        if index_file.is_file():
             self.tensor_names = set()
-            index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
-            index_name += ".index.json"
             logger.info(f"gguf: loading model weight map from '{index_name}'")
-            with open(self.dir_model / index_name, "r", encoding="utf-8") as f:
+            with open(index_file, "r", encoding="utf-8") as f:
                 index: dict[str, Any] = json.load(f)
                 weight_map = index.get("weight_map")
                 if weight_map is None or not isinstance(weight_map, dict):
@@ -145,6 +147,7 @@ class Model:
                 self.tensor_names.update(weight_map.keys())
         else:
             self.tensor_names = tensor_names_from_parts
+            weight_map = {}
 
         for part_name in self.part_names:
             logger.info(f"gguf: loading model part '{part_name}'")
@@ -171,9 +174,17 @@ class Model:
                             data = LazyTorchTensor.from_eager(data)
                     yield name, data
 
-        # only verify tensor name presence; it doesn't matter if they are not in the right files
-        if len(sym_diff := tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
-            raise ValueError(f"Mismatch between weight map and model parts for tensor names: {sym_diff}")
+        # verify tensor name presence and identify potentially missing files
+        if len(tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
+            missing = sorted(self.tensor_names.difference(tensor_names_from_parts))
+            extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
+            missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
+            if len(extra) == 0 and len(missing_files) > 0:
+                raise ValueError(f"Missing or incomplete model files: {missing_files}")
+            else:
+                raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
+                                 f"Missing tensors: {missing}\n"
+                                 f"Extra tensors: {extra}")
 
     def format_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
         if key not in gguf.MODEL_TENSORS[self.model_arch]:
@@ -2998,6 +3009,66 @@ class OlmoModel(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("OlmoeForCausalLM")
+class OlmoeModel(Model):
+    model_arch = gguf.MODEL_ARCH.OLMOE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_layer_norm_rms_eps(1e-5)
+        if (n_experts := self.hparams.get("num_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    # Copied from: Qwen2MoeModel
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # process the experts separately
+        if name.find("experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    # Copied from: Qwen2MoeModel
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @Model.register("JinaBertModel", "JinaBertForMaskedLM")
 class JinaBertV2Model(BertModel):
     model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
@@ -4009,6 +4080,36 @@ class ExaoneModel(Model):
         super().prepare_tensors()
 
 
+@Model.register("GraniteForCausalLM")
+class GraniteModel(LlamaModel):
+    """Conversion for IBM's GraniteForCausalLM"""
+    model_arch = gguf.MODEL_ARCH.GRANITE
+
+    def set_gguf_parameters(self):
+        """Granite uses standard llama parameters with the following differences:
+
+        - No head_dim support
+        - New multiplier params:
+            - attention_scale
+            - embedding_scale
+            - residual_scale
+        - logits_scaling
+        """
+        if head_dim := self.hparams.pop("head_dim", None):
+            logger.warning("Ignoring head_dim (%s) from config for Granite", head_dim)
+        super().set_gguf_parameters()
+        # NOTE: Convert _multiplier params to _scale params for naming
+        #   consistency
+        if attention_scale := self.hparams.get("attention_multiplier"):
+            self.gguf_writer.add_attention_scale(attention_scale)
+        if embedding_scale := self.hparams.get("embedding_multiplier"):
+            self.gguf_writer.add_embedding_scale(embedding_scale)
+        if residual_scale := self.hparams.get("residual_multiplier"):
+            self.gguf_writer.add_residual_scale(residual_scale)
+        if logits_scaling := self.hparams.get("logits_scaling"):
+            self.gguf_writer.add_logit_scale(logits_scaling)
+
+
 ###### CONVERSION LOGIC ######
 
 # tree of lazy tensors

docs/backend/SYCL.md

@@ -636,6 +636,14 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
   It's same for other projects including llama.cpp SYCL backend.
 
+- Meet issue: `Native API failed. Native API returns: -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -999 (UNKNOWN PI error)` or `failed to allocate SYCL0 buffer`
+
+  Device Memory is not enough.
+
+  |Reason|Solution|
+  |-|-|
+  |Default Context is too big. It leads to more memory usage.|Set `-c 8192` or smaller value.|
+  |Model is big and require more memory than device's.|Choose smaller quantized model, like Q5 -> Q4;<br>Use more than one devices to load model.|
 
 ### **GitHub contribution**:
 Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.

examples/imatrix/imatrix.cpp

@@ -572,6 +572,7 @@ int main(int argc, char ** argv) {
 
     params.n_ctx = 512;
     params.logits_all = true;
+    params.escape = false;
 
     if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
         return 1;

examples/llama-bench/llama-bench.cpp

@@ -439,6 +439,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 types.push_back(gt);
             }
+            if (invalid_param) {
+                break;
+            }
             params.type_k.insert(params.type_k.end(), types.begin(), types.end());
         } else if (arg == "-ctv" || arg == "--cache-type-v") {
             if (++i >= argc) {
@@ -455,6 +458,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 types.push_back(gt);
             }
+            if (invalid_param) {
+                break;
+            }
             params.type_v.insert(params.type_v.end(), types.begin(), types.end());
         } else if (arg == "-t" || arg == "--threads") {
             if (++i >= argc) {
@@ -520,6 +526,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 modes.push_back(mode);
             }
+            if (invalid_param) {
+                break;
+            }
             params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
         } else if (arg == "-mg" || arg == "--main-gpu") {
             if (++i >= argc) {

examples/perplexity/perplexity.cpp

@@ -1961,6 +1961,7 @@ int main(int argc, char ** argv) {
 
     params.n_ctx = 512;
     params.logits_all = true;
+    params.escape = false;
 
     if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
         return 1;

examples/server/README.md

@@ -87,7 +87,7 @@ The project is under active development, and we are [looking for feedback and co
 | `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16) |
 | `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16) |
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
-| `-np, --parallel N` | number of parallel sequences to decode (default: 1) |
+| `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env:  LLAMA_ARG_N_PARALLEL) |
 | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
 | `-nocb, --no-cont-batching` | disable continuous batching<br/>(env: LLAMA_ARG_NO_CONT_BATCHING) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing |
@@ -501,7 +501,7 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte
 
     See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
 
-    The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}`), similar to other OpenAI-inspired API providers.
+    The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name",  "type": "string" }, "date": { "title": "Date",  "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants",  "type": "string" } } } }`), similar to other OpenAI-inspired API providers.
 
     *Examples:*
 

examples/server/server.cpp

@@ -531,26 +531,38 @@ struct server_response {
 
     // add the id_task to the list of tasks waiting for response
     void add_waiting_task_id(int id_task) {
-        SRV_DBG("waiting for task id = %d\n", id_task);
+        SRV_DBG("add task %d to waiting list. current waiting = %d (before add)\n", id_task, (int) waiting_task_ids.size());
 
         std::unique_lock<std::mutex> lock(mutex_results);
         waiting_task_ids.insert(id_task);
     }
 
     void add_waiting_tasks(const std::vector<server_task> & tasks) {
-        for (const auto & t : tasks) {
-            add_waiting_task_id(t.id);
+        std::unique_lock<std::mutex> lock(mutex_results);
+
+        for (const auto & task : tasks) {
+            SRV_DBG("add task %d to waiting list. current waiting = %d (before add)\n", task.id, (int) waiting_task_ids.size());
+            waiting_task_ids.insert(task.id);
         }
     }
 
     // when the request is finished, we can remove task associated with it
     void remove_waiting_task_id(int id_task) {
-        SRV_DBG("task id = %d is done\n", id_task);
+        SRV_DBG("remove task %d from waiting list. current waiting = %d (before remove)\n", id_task, (int) waiting_task_ids.size());
 
         std::unique_lock<std::mutex> lock(mutex_results);
         waiting_task_ids.erase(id_task);
     }
 
+    void remove_waiting_task_ids(const std::unordered_set<int> & id_tasks) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+
+        for (const auto & id_task : id_tasks) {
+            SRV_DBG("remove task %d from waiting list. current waiting = %d (before remove)\n", id_task, (int) waiting_task_ids.size());
+            waiting_task_ids.erase(id_task);
+        }
+    }
+
     // This function blocks the thread until there is a response for one of the id_tasks
     server_task_result recv(const std::unordered_set<int> & id_tasks) {
         while (true) {
@@ -2254,14 +2266,6 @@ static void log_server_request(const httplib::Request & req, const httplib::Resp
         return;
     }
 
-    //LOG_INFO("request", {
-    //    {"remote_addr", req.remote_addr},
-    //    {"remote_port", req.remote_port},
-    //    {"status",      res.status},
-    //    {"method",      req.method},
-    //    {"path",        req.path},
-    //    {"params",      req.params},
-    //});
     LOG_INF("request: %s %s %s %d\n", req.method.c_str(), req.path.c_str(), req.remote_addr.c_str(), res.status);
 
     LOG_DBG("request:  %s\n", req.body.c_str());
@@ -2318,12 +2322,12 @@ int main(int argc, char ** argv) {
     std::unique_ptr<httplib::Server> svr;
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
     if (params.ssl_file_key != "" && params.ssl_file_cert != "") {
-        LOG_INFO("Running with SSL", {{"key", params.ssl_file_key}, {"cert", params.ssl_file_cert}});
+        LOG_INF("Running with SSL: key = %s, cert = %s\n", params.ssl_file_key.c_str(), params.ssl_file_cert.c_str());
         svr.reset(
             new httplib::SSLServer(params.ssl_file_cert.c_str(), params.ssl_file_key.c_str())
         );
     } else {
-        LOG_INFO("Running without SSL", {});
+        LOG_INF("Running without SSL\n");
         svr.reset(new httplib::Server());
     }
 #else
@@ -2782,6 +2786,8 @@ int main(int argc, char ** argv) {
             }, [&](const json & error_data) {
                 res_error(res, error_data);
             });
+
+            ctx_server.queue_results.remove_waiting_task_ids(task_ids);
         } else {
             const auto chunked_content_provider = [task_ids, &ctx_server](size_t, httplib::DataSink & sink) {
                 ctx_server.receive_cmpl_results_stream(task_ids, [&](const server_task_result & result) -> bool {
@@ -2792,7 +2798,12 @@ int main(int argc, char ** argv) {
                 sink.done();
                 return false;
             };
-            res.set_chunked_content_provider("text/event-stream", chunked_content_provider);
+
+            auto on_complete = [task_ids, &ctx_server] (bool) {
+                ctx_server.queue_results.remove_waiting_task_ids(task_ids);
+            };
+
+            res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
         }
     };
 
@@ -2831,6 +2842,8 @@ int main(int argc, char ** argv) {
             }, [&](const json & error_data) {
                 res_error(res, error_data);
             });
+
+            ctx_server.queue_results.remove_waiting_task_ids(task_ids);
         } else {
             const auto chunked_content_provider = [task_ids, &ctx_server, completion_id](size_t, httplib::DataSink & sink) {
                 ctx_server.receive_cmpl_results_stream(task_ids, [&](const server_task_result & result) -> bool {
@@ -2852,7 +2865,12 @@ int main(int argc, char ** argv) {
                 sink.done();
                 return true;
             };
-            res.set_chunked_content_provider("text/event-stream", chunked_content_provider);
+
+            auto on_complete = [task_ids, &ctx_server] (bool) {
+                ctx_server.queue_results.remove_waiting_task_ids(task_ids);
+            };
+
+            res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
         }
     };
 
@@ -2961,6 +2979,8 @@ int main(int argc, char ** argv) {
                 res_error(res, error_data);
                 error = true;
             });
+
+            ctx_server.queue_results.remove_waiting_task_ids(task_ids);
         }
 
         if (error) {
@@ -3108,7 +3128,6 @@ int main(int argc, char ** argv) {
     std::thread t([&]() { svr->listen_after_bind(); });
     svr->wait_until_ready();
 
-    //LOG_INFO("HTTP server is listening", log_data);
     LOG_INF("%s: HTTP server is listening, hostname: %s, port: %d, http threads: %d\n", __func__, params.hostname.c_str(), params.port, params.n_threads_http);
 
     // load the model

examples/server/utils.hpp

@@ -331,6 +331,9 @@ static json oaicompat_completion_params_parse(
         std::string response_type = json_value(response_format, "type", std::string());
         if (response_type == "json_object") {
             llama_params["json_schema"] = json_value(response_format, "schema", json::object());
+        } else if (response_type == "json_schema") {
+            json json_schema = json_value(response_format, "json_schema", json::object());
+            llama_params["json_schema"] = json_value(json_schema, "schema", json::object());
         } else if (!response_type.empty() && response_type != "text") {
             throw std::runtime_error("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type);
         }

examples/sycl/run-llama2.sh

@@ -11,16 +11,17 @@ source /opt/intel/oneapi/setvars.sh
 #ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
 
 INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
-MODEL_FILE=llama-2-7b.Q4_0.gguf
+MODEL_FILE=models/llama-2-7b.Q4_0.gguf
 NGL=33
+CONEXT=8192
 
 if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONEXT} -mg $GGML_SYCL_DEVICE -sm none
 
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONEXT}
 fi

ggml/CMakeLists.txt

@@ -56,6 +56,15 @@ else()
     set(GGML_NATIVE_DEFAULT ON)
 endif()
 
+# defaults
+if (NOT GGML_LLAMAFILE_DEFAULT)
+    set(GGML_LLAMAFILE_DEFAULT OFF)
+endif()
+
+if (NOT GGML_CUDA_GRAPHS_DEFAULT)
+    set(GGML_CUDA_GRAPHS_DEFAULT OFF)
+endif()
+
 # general
 option(GGML_STATIC "ggml: static link libraries"         OFF)
 option(GGML_NATIVE "ggml: enable -march=native flag"     ${GGML_NATIVE_DEFAULT})
@@ -110,7 +119,7 @@ option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"
 option(GGML_BLAS                            "ggml: use BLAS"                                  ${GGML_BLAS_DEFAULT})
 set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING
                                             "ggml: BLAS library vendor")
-option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"                             OFF)
+option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"                             ${GGML_LLAMAFILE_DEFAULT})
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
@@ -127,7 +136,7 @@ set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
 option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"                OFF)
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
-option(GGML_CUDA_USE_GRAPHS                 "ggml: use CUDA graphs (llama.cpp only)"          OFF)
+option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
 
 option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)

ggml/src/CMakeLists.txt

@@ -329,7 +329,7 @@ if (GGML_CUDA)
         add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 
-        if (GGML_CUDA_USE_GRAPHS)
+        if (GGML_CUDA_GRAPHS)
             add_compile_definitions(GGML_CUDA_USE_GRAPHS)
         endif()
 
@@ -1341,7 +1341,7 @@ list(APPEND GGML_EXTRA_LIBS_PRIVATE Threads::Threads)
 find_library(MATH_LIBRARY m)
 if (MATH_LIBRARY)
     if (NOT WIN32 OR NOT GGML_SYCL)
-        target_link_libraries(ggml PRIVATE ${MATH_LIBRARY})
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE m)
     endif()
 endif()
 

ggml/src/ggml-aarch64.c

@@ -4,6 +4,7 @@
 
 #include "ggml-quants.h"
 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
 
 #include <math.h>
 #include <string.h>

ggml/src/ggml-cpu-impl.h

@@ -0,0 +1,614 @@
+#pragma once
+
+// GGML CPU internal header
+
+#include "ggml.h"
+#include "ggml-impl.h"
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
+//#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(_MSC_VER)
+
+#define m512bh(p) p
+#define m512i(p) p
+
+#else
+
+#define m512bh(p) (__m512bh)(p)
+#define m512i(p) (__m512i)(p)
+
+#endif
+
+/**
+ * Converts brain16 to float32.
+ *
+ * The bfloat16 floating point format has the following structure:
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───┐
+ *     0b0000000000000000 brain16
+ *
+ * Since bf16 has the same number of exponent bits as a 32bit float,
+ * encoding and decoding numbers becomes relatively straightforward.
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───────────────────┐
+ *     0b00000000000000000000000000000000 IEEE binary32
+ *
+ * For comparison, the standard fp16 format has fewer exponent bits.
+ *
+ *       ┌sign
+ *       │
+ *       │  ┌exponent
+ *       │  │
+ *       │  │    ┌mantissa
+ *       │  │    │
+ *       │┌─┴─┐┌─┴──────┐
+ *     0b0000000000000000 IEEE binary16
+ *
+ * @see IEEE 754-2008
+ */
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+/**
+ * Converts float32 to brain16.
+ *
+ * This is binary identical with Google Brain float conversion.
+ * Floats shall round to nearest even, and NANs shall be quiet.
+ * Subnormals aren't flushed to zero, except perhaps when used.
+ * This code should vectorize nicely if using modern compilers.
+ */
+static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
+    ggml_bf16_t h;
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.f = s;
+    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
+        h.bits = (u.i >> 16) | 64; /* force to quiet */
+        return h;
+    }
+    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
+    return h;
+}
+
+#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
+#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#endif
+
+// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
+#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#ifndef __SSSE3__
+#define __SSSE3__
+#endif
+#endif
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+#include <sys/prctl.h>
+#endif
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+#if defined(__ARM_NEON)
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#ifdef _MSC_VER
+
+typedef uint16_t ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+
+#else
+
+typedef __fp16 ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+
+#endif // _MSC_VER
+
+#if !defined(__aarch64__)
+
+// 32-bit ARM compatibility
+
+// vaddlvq_s16
+// vpaddq_s16
+// vpaddq_s32
+// vaddvq_s32
+// vaddvq_f32
+// vmaxvq_f32
+// vcvtnq_s32_f32
+// vzip1_u8
+// vzip2_u8
+
+inline static int32_t vaddlvq_s16(int16x8_t v) {
+    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
+    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
+    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
+    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
+    return vcombine_s32(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+
+inline static float vaddvq_f32(float32x4_t v) {
+    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
+}
+
+inline static float vmaxvq_f32(float32x4_t v) {
+    return
+        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
+            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
+}
+
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
+}
+
+inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
+}
+
+inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
+}
+
+// vld1q_s16_x2
+// vld1q_u8_x2
+// vld1q_u8_x4
+// vld1q_s8_x2
+// vld1q_s8_x4
+// TODO: double-check these work correctly
+
+typedef struct ggml_int16x8x2_t {
+    int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
+    ggml_int16x8x2_t res;
+
+    res.val[0] = vld1q_s16(ptr + 0);
+    res.val[1] = vld1q_s16(ptr + 8);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
+    ggml_uint8x16x2_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
+    ggml_uint8x16x4_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+    res.val[2] = vld1q_u8(ptr + 32);
+    res.val[3] = vld1q_u8(ptr + 48);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x2_t {
+    int8x16_t val[2];
+} ggml_int8x16x2_t;
+
+inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
+    ggml_int8x16x2_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+    int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
+    ggml_int8x16x4_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+    res.val[2] = vld1q_s8(ptr + 32);
+    res.val[3] = vld1q_s8(ptr + 48);
+
+    return res;
+}
+
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+// NOTE: not tested
+inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+#else
+
+#define ggml_int16x8x2_t  int16x8x2_t
+#define ggml_uint8x16x2_t uint8x16x2_t
+#define ggml_uint8x16x4_t uint8x16x4_t
+#define ggml_int8x16x2_t  int8x16x2_t
+#define ggml_int8x16x4_t  int8x16x4_t
+
+#define ggml_vld1q_s16_x2 vld1q_s16_x2
+#define ggml_vld1q_u8_x2  vld1q_u8_x2
+#define ggml_vld1q_u8_x4  vld1q_u8_x4
+#define ggml_vld1q_s8_x2  vld1q_s8_x2
+#define ggml_vld1q_s8_x4  vld1q_s8_x4
+#define ggml_vqtbl1q_s8   vqtbl1q_s8
+#define ggml_vqtbl1q_u8   vqtbl1q_u8
+
+#endif // !defined(__aarch64__)
+
+#if !defined(__ARM_FEATURE_DOTPROD)
+
+inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
+    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+
+    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
+}
+
+#else
+
+#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
+
+#endif // !defined(__ARM_FEATURE_DOTPROD)
+
+#endif // defined(__ARM_NEON)
+
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    ggml_fp16_internal_t tmp;
+    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+    return (float)tmp;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    ggml_fp16_t res;
+    ggml_fp16_internal_t tmp = f;
+    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+    return res;
+}
+
+#else
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#if defined(__loongarch64)
+#if defined(__loongarch_asx)
+#include <lasxintrin.h>
+#endif
+#if defined(__loongarch_sx)
+#include <lsxintrin.h>
+#endif
+#endif
+
+#if defined(__loongarch_asx)
+
+typedef union {
+    int32_t i;
+    float f;
+} ft_union;
+
+/* float type data load instructions */
+static __m128 __lsx_vreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
+}
+
+static __m256 __lasx_xvreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
+}
+#endif
+
+#ifdef __F16C__
+
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
+#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
+
+#elif defined(__POWER9_VECTOR__)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+/* the inline asm below is about 12% faster than the lookup method */
+#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    register float f;
+    register double d;
+    __asm__(
+        "mtfprd %0,%2\n"
+        "xscvhpdp %0,%0\n"
+        "frsp %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=f"(f):
+        /* in */   "r"(h));
+    return f;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    register double d;
+    register ggml_fp16_t r;
+    __asm__( /* xscvdphp can work on double or single precision */
+        "xscvdphp %0,%2\n"
+        "mffprd %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=r"(r):
+        /* in */   "f"(f));
+    return r;
+}
+
+#else
+
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
+
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
+}
+
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+#endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // __F16C__
+
+#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
+
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32)
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(GGML_FP32_TO_FP16)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-impl.h

@@ -1,15 +1,17 @@
 #pragma once
 
-#include "ggml.h"
-
 // GGML internal header
 
+#include "ggml.h"
+
 #include <assert.h>
 #include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
-#include <stddef.h>
 #include <stdbool.h>
-#include <string.h> // memcpy
-#include <math.h>   // fabsf
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
 
 #undef MIN
 #undef MAX
@@ -17,96 +19,6 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
-#if defined(_MSC_VER)
-
-#define m512bh(p) p
-#define m512i(p) p
-
-#else
-
-#define m512bh(p) (__m512bh)(p)
-#define m512i(p) (__m512i)(p)
-
-#endif
-
-/**
- * Converts brain16 to float32.
- *
- * The bfloat16 floating point format has the following structure:
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───┐
- *     0b0000000000000000 brain16
- *
- * Since bf16 has the same number of exponent bits as a 32bit float,
- * encoding and decoding numbers becomes relatively straightforward.
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───────────────────┐
- *     0b00000000000000000000000000000000 IEEE binary32
- *
- * For comparison, the standard fp16 format has fewer exponent bits.
- *
- *       ┌sign
- *       │
- *       │  ┌exponent
- *       │  │
- *       │  │    ┌mantissa
- *       │  │    │
- *       │┌─┴─┐┌─┴──────┐
- *     0b0000000000000000 IEEE binary16
- *
- * @see IEEE 754-2008
- */
-static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.i = (uint32_t)h.bits << 16;
-    return u.f;
-}
-
-/**
- * Converts float32 to brain16.
- *
- * This is binary identical with Google Brain float conversion.
- * Floats shall round to nearest even, and NANs shall be quiet.
- * Subnormals aren't flushed to zero, except perhaps when used.
- * This code should vectorize nicely if using modern compilers.
- */
-static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
-    ggml_bf16_t h;
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.f = s;
-    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
-        h.bits = (u.i >> 16) | 64; /* force to quiet */
-        return h;
-    }
-    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
-    return h;
-}
-
-#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
-#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
 // static_assert should be a #define, but if it's not,
 // fall back to the _Static_assert C11 keyword.
 // if C99 - static_assert is noop
@@ -121,520 +33,6 @@ extern "C" {
 #endif
 #endif
 
-// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
-#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __FMA__
-#define __FMA__
-#endif
-#ifndef __F16C__
-#define __F16C__
-#endif
-#endif
-
-// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
-#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __SSE3__
-#define __SSE3__
-#endif
-#ifndef __SSSE3__
-#define __SSSE3__
-#endif
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-#include <arm_sve.h>
-#include <sys/prctl.h>
-#endif
-
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-#if defined(__ARM_NEON)
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-#ifdef _MSC_VER
-
-typedef uint16_t ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
-
-#else
-
-typedef __fp16 ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
-
-#endif // _MSC_VER
-
-#if !defined(__aarch64__)
-
-// 32-bit ARM compatibility
-
-// vaddlvq_s16
-// vpaddq_s16
-// vpaddq_s32
-// vaddvq_s32
-// vaddvq_f32
-// vmaxvq_f32
-// vcvtnq_s32_f32
-// vzip1_u8
-// vzip2_u8
-
-inline static int32_t vaddlvq_s16(int16x8_t v) {
-    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
-    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
-}
-
-inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
-    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
-    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
-    return vcombine_s16(a0, b0);
-}
-
-inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
-    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
-    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
-    return vcombine_s32(a0, b0);
-}
-
-inline static int32_t vaddvq_s32(int32x4_t v) {
-    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
-}
-
-inline static float vaddvq_f32(float32x4_t v) {
-    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
-}
-
-inline static float vmaxvq_f32(float32x4_t v) {
-    return
-        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
-            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
-}
-
-inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
-    int32x4_t res;
-
-    res[0] = roundf(vgetq_lane_f32(v, 0));
-    res[1] = roundf(vgetq_lane_f32(v, 1));
-    res[2] = roundf(vgetq_lane_f32(v, 2));
-    res[3] = roundf(vgetq_lane_f32(v, 3));
-
-    return res;
-}
-
-inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[0]; res[1] = b[0];
-    res[2] = a[1]; res[3] = b[1];
-    res[4] = a[2]; res[5] = b[2];
-    res[6] = a[3]; res[7] = b[3];
-
-    return res;
-}
-
-inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[4]; res[1] = b[4];
-    res[2] = a[5]; res[3] = b[5];
-    res[4] = a[6]; res[5] = b[6];
-    res[6] = a[7]; res[7] = b[7];
-
-    return res;
-}
-
-// vld1q_s16_x2
-// vld1q_u8_x2
-// vld1q_u8_x4
-// vld1q_s8_x2
-// vld1q_s8_x4
-// TODO: double-check these work correctly
-
-typedef struct ggml_int16x8x2_t {
-    int16x8_t val[2];
-} ggml_int16x8x2_t;
-
-inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
-    ggml_int16x8x2_t res;
-
-    res.val[0] = vld1q_s16(ptr + 0);
-    res.val[1] = vld1q_s16(ptr + 8);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x2_t {
-    uint8x16_t val[2];
-} ggml_uint8x16x2_t;
-
-inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
-    ggml_uint8x16x2_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x4_t {
-    uint8x16_t val[4];
-} ggml_uint8x16x4_t;
-
-inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
-    ggml_uint8x16x4_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-    res.val[2] = vld1q_u8(ptr + 32);
-    res.val[3] = vld1q_u8(ptr + 48);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x2_t {
-    int8x16_t val[2];
-} ggml_int8x16x2_t;
-
-inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
-    ggml_int8x16x2_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x4_t {
-    int8x16_t val[4];
-} ggml_int8x16x4_t;
-
-inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
-    ggml_int8x16x4_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-    res.val[2] = vld1q_s8(ptr + 32);
-    res.val[3] = vld1q_s8(ptr + 48);
-
-    return res;
-}
-
-// NOTE: not tested
-inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
-    int8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-// NOTE: not tested
-inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
-    uint8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-#else
-
-#define ggml_int16x8x2_t  int16x8x2_t
-#define ggml_uint8x16x2_t uint8x16x2_t
-#define ggml_uint8x16x4_t uint8x16x4_t
-#define ggml_int8x16x2_t  int8x16x2_t
-#define ggml_int8x16x4_t  int8x16x4_t
-
-#define ggml_vld1q_s16_x2 vld1q_s16_x2
-#define ggml_vld1q_u8_x2  vld1q_u8_x2
-#define ggml_vld1q_u8_x4  vld1q_u8_x4
-#define ggml_vld1q_s8_x2  vld1q_s8_x2
-#define ggml_vld1q_s8_x4  vld1q_s8_x4
-#define ggml_vqtbl1q_s8   vqtbl1q_s8
-#define ggml_vqtbl1q_u8   vqtbl1q_u8
-
-#endif // !defined(__aarch64__)
-
-#if !defined(__ARM_FEATURE_DOTPROD)
-
-inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
-    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
-    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
-
-    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
-}
-
-#else
-
-#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
-
-#endif // !defined(__ARM_FEATURE_DOTPROD)
-
-#endif // defined(__ARM_NEON)
-
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
-
-#else
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#ifdef __POWER9_VECTOR__
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#if defined(__loongarch64)
-#if defined(__loongarch_asx)
-#include <lasxintrin.h>
-#endif
-#if defined(__loongarch_sx)
-#include <lsxintrin.h>
-#endif
-#endif
-
-#if defined(__loongarch_asx)
-
-typedef union {
-    int32_t i;
-    float f;
-} ft_union;
-
-/* float type data load instructions */
-static __m128 __lsx_vreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
-}
-
-static __m256 __lasx_xvreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
-}
-#endif
-
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
-
-#ifdef __ARM_FEATURE_SVE
-#include <arm_sve.h>
-#endif // __ARM_FEATURE_SVE
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-#endif
-
-enum ggml_cgraph_eval_order {
-    GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
-    GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
-    GGML_CGRAPH_EVAL_ORDER_COUNT
-};
-
 // bitset
 
 typedef uint32_t ggml_bitset_t;
@@ -761,6 +159,12 @@ static size_t ggml_hash_find_or_insert(struct ggml_hash_set * hash_set, struct g
 
 // computation graph
 
+enum ggml_cgraph_eval_order {
+    GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
+    GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
+    GGML_CGRAPH_EVAL_ORDER_COUNT
+};
+
 struct ggml_cgraph {
     int size;
     int n_nodes;

ggml/src/ggml-quants.c

@@ -3,6 +3,7 @@
 
 #include "ggml-quants.h"
 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
 
 
 #include <math.h>
@@ -230,6 +231,12 @@ static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
 
     return _mm_packus_epi16( bytes1, bytes2);
 }
+
+static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
+    const __m128i ax = _mm_sign_epi8(x, x);
+    const __m128i sy = _mm_sign_epi8(y, x);
+    return _mm_maddubs_epi16(ax, sy);
+}
 #endif
 #elif defined(__SSSE3__)
 // horizontally add 4x4 floats
@@ -4206,37 +4213,37 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
 
     sumf = hsum_float_8(acc);
 #elif defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
+    const __m128i mone = _mm_set1_epi16(1);
 
-    // Main loop
-    for (; ib < nb; ++ib) {
-        // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
 
-        const __m128i lowMask = _mm_set1_epi8(0xF);
-        const __m128i off = _mm_set1_epi8(8);
-
-        const __m128i tmp = _mm_loadu_si128((const __m128i *)x[ib].qs);
-
-        __m128i bx_0 = _mm_and_si128(lowMask, tmp);
-        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
-        bx_0 = _mm_sub_epi8(bx_0, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
-
-        bx_0 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4));
-        by_0 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
-        bx_0 = _mm_sub_epi8(bx_0, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx_0, by_0);
-
-        // Convert int32_t to float
-        __m256 p = _mm256_cvtepi32_ps(MM256_SET_M128I(i32_0, i32_1));
-
-        // Apply the scale, and accumulate
-        acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc);
+        const __m128i q4b_1_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_1), _mm_set1_epi8(8));
+        const __m128i q4b_1_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_1, 4)), _mm_set1_epi8(8));
+        const __m128i q4b_2_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_2), _mm_set1_epi8(8));
+        const __m128i q4b_2_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_2, 4)), _mm_set1_epi8(8));
+        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+        const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
+        const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
+        const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
+        const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
+        accum1 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_1_1, p_1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_2_1, p_2_0))), accum2);
     }
 
-    sumf = hsum_float_8(acc);
+    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
 #elif defined(__SSSE3__)
     // set constants
     const __m128i lowMask = _mm_set1_epi8(0xF);
@@ -11819,15 +11826,6 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
 #endif
 }
 
-
-#if defined(__AVX__)
-static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
-    const __m128i ax = _mm_sign_epi8(x, x);
-    const __m128i sy = _mm_sign_epi8(y, x);
-    return _mm_maddubs_epi16(ax, sy);
-}
-#endif
-
 #if defined(__AVX2__)
 static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
     const __m256i ax = _mm256_sign_epi8(x, x);

ggml/src/ggml.c

@@ -2,6 +2,7 @@
 #define _USE_MATH_DEFINES // For M_PI on MSVC
 
 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
 #include "ggml.h"
 #include "ggml-aarch64.h"
@@ -2012,10 +2013,11 @@ struct ggml_threadpool {
     // these are atomic as an annotation for thread-sanitizer
     atomic_bool stop;         // Used for stopping the threadpool altogether
     atomic_bool pause;        // Used for pausing the threadpool or individual threads
+    atomic_bool abort;        // Used for aborting processing of a graph
 
     struct ggml_compute_state * workers;   // per thread state
     int          n_threads_max; // number of threads in the pool
-    int          n_threads_cur; // number of threads used in the current graph
+    atomic_int   n_threads_cur; // number of threads used in the current graph
 
     int32_t      prio;        // Scheduling priority
     uint32_t     poll;        // Polling level (0 - no polling)
@@ -3177,41 +3179,36 @@ inline static void ggml_critical_section_start(void) {
     }
 }
 
+static void ggml_barrier(struct ggml_threadpool * tp) {
+    int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
+    if (n_threads == 1) {
+        return;
+    }
+
 #ifdef GGML_USE_OPENMP
-static void ggml_barrier(struct ggml_threadpool * threadpool) {
-    if (threadpool->n_threads_cur == 1) {
-        return;
-    }
-
     #pragma omp barrier
-}
 #else
-static void ggml_barrier(struct ggml_threadpool * threadpool) {
-    if (threadpool->n_threads_cur == 1) {
-        return;
-    }
+    int n_passed = atomic_load_explicit(&tp->n_barrier_passed, memory_order_relaxed);
 
-    atomic_int * n_barrier = &threadpool->n_barrier;
-    atomic_int * n_barrier_passed = &threadpool->n_barrier_passed;
+    // enter barrier (full seq-cst fence)
+    int n_barrier = atomic_fetch_add_explicit(&tp->n_barrier, 1, memory_order_seq_cst);
 
-    int n_threads = threadpool->n_threads_cur;
-    int passed_old = atomic_load_explicit(n_barrier_passed, memory_order_relaxed);
-
-    if (atomic_fetch_add(n_barrier, 1) == n_threads - 1) {
+    int last = 0;
+    if (n_barrier == (n_threads - 1)) {
         // last thread
-        atomic_store(n_barrier, 0);
-        atomic_fetch_add_explicit(n_barrier_passed, 1, memory_order_relaxed);
+        atomic_store_explicit(&tp->n_barrier, 0, memory_order_relaxed);
+        last = 1;
     } else {
         // wait for other threads
-        while (true) {
-            if (atomic_load_explicit(n_barrier_passed, memory_order_relaxed) != passed_old) {
-                return;
-            }
+        while (atomic_load_explicit(&tp->n_barrier_passed, memory_order_relaxed) == n_passed) {
             ggml_thread_cpu_relax();
         }
     }
-}
+
+    // exit barrier (full seq-cst fence)
+    atomic_fetch_add_explicit(&tp->n_barrier_passed, last, memory_order_seq_cst);
 #endif
+}
 
 // TODO: make this somehow automatically executed
 //       some sort of "sentry" mechanism
@@ -19932,34 +19929,33 @@ struct ggml_cplan ggml_graph_plan(
 
 static thread_ret_t ggml_graph_compute_thread(void * data) {
     struct ggml_compute_state * state = (struct ggml_compute_state *) data;
+    struct ggml_threadpool    * tp    = state->threadpool;
 
-    const struct ggml_cgraph * cgraph = state->threadpool->cgraph;
-    const struct ggml_cplan  * cplan  = state->threadpool->cplan;
+    const struct ggml_cgraph * cgraph = tp->cgraph;
+    const struct ggml_cplan  * cplan  = tp->cplan;
 
     set_numa_thread_affinity(state->ith);
 
     struct ggml_compute_params params = {
         /*.ith       =*/ state->ith,
-        /*.nth       =*/ state->threadpool->n_threads_cur,
+        /*.nth       =*/ atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed),
         /*.wsize     =*/ cplan->work_size,
         /*.wdata     =*/ cplan->work_data,
-        /*.threadpool=*/ state->threadpool,
+        /*.threadpool=*/ tp,
     };
 
-    for (int node_n = 0; node_n < cgraph->n_nodes; node_n++) {
+    for (int node_n = 0; node_n < cgraph->n_nodes && !tp->abort; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
 
         ggml_compute_forward(&params, node);
 
-        if (state->ith == 0 && cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) {
-            state->threadpool->ec = GGML_STATUS_ABORTED;
+        if (state->ith == 0 && cplan->abort_callback &&
+                cplan->abort_callback(cplan->abort_callback_data)) {
+            tp->abort = true;
+            tp->ec    = GGML_STATUS_ABORTED;
         }
 
         ggml_barrier(state->threadpool);
-
-        if (state->threadpool->ec != GGML_STATUS_SUCCESS) {
-            break;
-        }
     }
 
     return 0;
@@ -19967,7 +19963,15 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
 
 #ifndef GGML_USE_OPENMP
 
-static inline bool ggml_graph_compute_ready(struct ggml_compute_state * state) {
+// check if thread is active
+static inline bool ggml_graph_compute_thread_active(struct ggml_compute_state * state) {
+    struct ggml_threadpool * threadpool = state->threadpool;
+    int n_threads = atomic_load_explicit(&threadpool->n_threads_cur, memory_order_relaxed);
+    return (state->ith < n_threads);
+}
+
+// check if thread is ready to proceed (exit from polling or sleeping)
+static inline bool ggml_graph_compute_thread_ready(struct ggml_compute_state * state) {
     struct ggml_threadpool * threadpool = state->threadpool;
 
     if (state->pending || threadpool->stop || threadpool->pause) { return true; }
@@ -19975,21 +19979,34 @@ static inline bool ggml_graph_compute_ready(struct ggml_compute_state * state) {
     // check for new graph/work
     int new_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed);
     if (new_graph != state->last_graph) {
-        state->pending    = (state->ith < threadpool->n_threads_cur);
+        state->pending    = ggml_graph_compute_thread_active(state);
         state->last_graph = new_graph;
     }
 
     return state->pending;
 }
 
+// sync thread state after polling
+static inline void ggml_graph_compute_thread_sync(struct ggml_compute_state * state) {
+    struct ggml_threadpool * threadpool = state->threadpool;
+    // this should just be atomic_thread_fence(seq_cst) but it confuses thread-sanitizer
+    // so instead we just use a dummy read-modify-write
+    atomic_fetch_add_explicit(&threadpool->n_graph, 0, memory_order_seq_cst);
+}
+
 static inline bool ggml_graph_compute_poll_for_work(struct ggml_compute_state * state) {
     struct ggml_threadpool * threadpool = state->threadpool;
 
+    // Skip polling for unused threads
+    if (!ggml_graph_compute_thread_active(state)) {
+        return state->pending;
+    }
+
     // This seems to make 0 ... 100 a decent range for polling level across modern processors.
     // Perhaps, we can adjust it dynamically based on load and things.
     const uint64_t n_rounds = 1024UL * 128 * threadpool->poll;
 
-    for (uint64_t i=0; !ggml_graph_compute_ready(state) && i<n_rounds; i++) {
+    for (uint64_t i=0; !ggml_graph_compute_thread_ready(state) && i < n_rounds; i++) {
         // No new work. Keep polling.
         ggml_thread_cpu_relax();
     }
@@ -20001,13 +20018,14 @@ static inline bool ggml_graph_compute_check_for_work(struct ggml_compute_state *
     struct ggml_threadpool * threadpool = state->threadpool;
 
     if (ggml_graph_compute_poll_for_work(state)) {
+        ggml_graph_compute_thread_sync(state);
         return state->pending;
     }
 
     ggml_mutex_lock_shared(&threadpool->mutex);
-    while (!ggml_graph_compute_ready(state)) {
+    while (!ggml_graph_compute_thread_ready(state)) {
         // No new work. Wait for the signal.
-        GGML_PRINT_DEBUG("thread #%d waiting for work\n", state->ith);
+        GGML_PRINT_DEBUG("thread #%d waiting for work (sleeping)\n", state->ith);
         ggml_cond_wait(&threadpool->cond, &threadpool->mutex);
     }
     ggml_mutex_unlock_shared(&threadpool->mutex);
@@ -20054,13 +20072,20 @@ static thread_ret_t ggml_graph_compute_secondary_thread(void* data) {
 }
 
 // Start processing new graph
-static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool)
+static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int n_threads)
 {
-    // always take the mutex here because the worker threads are doing hybrid poll/wait
+    // Always take the mutex here because the worker threads are doing hybrid poll/wait
 
     ggml_mutex_lock(&threadpool->mutex);
 
-    atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_relaxed);
+    GGML_PRINT_DEBUG("threadpool: n_threads_cur %d n_threads %d\n", threadpool->n_threads_cur, n_threads);
+
+    // Update the number of active threads
+    atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
+
+    // Indicate the graph is ready to be processed
+    // We need the full seq-cst fence here because of the polling threads (used in thread_sync)
+    atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_seq_cst);
 
     if (threadpool->pause) {
        // Update main thread prio and affinity to match the threadpool settings
@@ -20119,6 +20144,7 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
         threadpool->current_chunk    = 0;
         threadpool->stop             = false;
         threadpool->pause            = tpp->paused;
+        threadpool->abort            = false;
         threadpool->workers          = NULL;
         threadpool->n_threads_max    = tpp->n_threads;
         threadpool->n_threads_cur    = tpp->n_threads;
@@ -20194,15 +20220,11 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
         // No worker threads should be accessing the parameters below at this stage
         threadpool->cgraph           = cgraph;
         threadpool->cplan            = cplan;
-        threadpool->n_threads_cur    = n_threads;
         threadpool->current_chunk    = 0;
+        threadpool->abort            = false;
         threadpool->ec               = GGML_STATUS_SUCCESS;
     }
 
-    if (n_threads > threadpool->n_threads_max) {
-        GGML_PRINT("WARNING: cplan is requesting more threads than the threadpool contains. Expect a bad time!\n");
-    }
-
 #ifdef GGML_USE_OPENMP
     if (n_threads > 1) {
         #pragma omp parallel num_threads(n_threads)
@@ -20211,17 +20233,23 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
             {
                 // update the number of threads from the actual number of threads that we got from OpenMP
                 n_threads = omp_get_num_threads();
-                threadpool->n_threads_cur = n_threads;
+                atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
             }
 
             ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]);
         }
     } else {
+        atomic_store_explicit(&threadpool->n_threads_cur, 1, memory_order_relaxed);
         ggml_graph_compute_thread(&threadpool->workers[0]);
     }
 #else
+    if (n_threads > threadpool->n_threads_max) {
+        GGML_PRINT("WARNING: cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads_max);
+        n_threads = threadpool->n_threads_max;
+    }
+
     // Kick all threads to start the new graph
-    ggml_graph_compute_kickoff(threadpool);
+    ggml_graph_compute_kickoff(threadpool, n_threads);
 
     // This is a work thread too
     ggml_graph_compute_thread(&threadpool->workers[0]);

ggml/src/llamafile/sgemm.cpp

@@ -50,6 +50,7 @@
 
 #include "sgemm.h"
 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
 
 #ifdef _MSC_VER
@@ -235,6 +236,14 @@ template <> inline __m512 load(const ggml_fp16_t *p) {
 }
 #endif // __AVX512F__
 
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// CONSTANTS
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // FLOATING POINT MATRIX MULTIPLICATION
 
@@ -933,6 +942,20 @@ class tinyBLAS_Q0_AVX {
         return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
     }
 
+    inline __m256i load(const block_iq4_nl *b) {
+        return MM256_SET_M128I(load1(b), load0(b));
+    }
+
+    inline __m128i load0(const block_iq4_nl *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), x));
+    }
+
+    inline __m128i load1(const block_iq4_nl *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)));
+    }
+
     inline __m256 updot(__m256i u, __m256i s) {
         __m256i res;
 #if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
@@ -1159,6 +1182,22 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
 #endif
     }
 
+    case GGML_TYPE_IQ4_NL: {
+        if (Btype != GGML_TYPE_Q8_0)
+            return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_iq4_nl, block_q8_0, float> tb{
+            k, (const block_iq4_nl *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
     default:
         return false;
     }

gguf-py/gguf/constants.py

@@ -97,6 +97,8 @@ class Keys:
         RESCALE_EVERY_N_LAYERS            = "{arch}.rescale_every_n_layers"
         TIME_MIX_EXTRA_DIM                = "{arch}.time_mix_extra_dim"
         TIME_DECAY_EXTRA_DIM              = "{arch}.time_decay_extra_dim"
+        RESIDUAL_SCALE                    = "{arch}.residual_scale"
+        EMBEDDING_SCALE                   = "{arch}.embedding_scale"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"
@@ -112,6 +114,7 @@ class Keys:
         KV_LORA_RANK      = "{arch}.attention.kv_lora_rank"
         REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count"
         SLIDING_WINDOW    = "{arch}.attention.sliding_window"
+        SCALE             = "{arch}.attention.scale"
 
     class Rope:
         DIMENSION_COUNT         = "{arch}.rope.dimension_count"
@@ -220,6 +223,7 @@ class MODEL_ARCH(IntEnum):
     COMMAND_R    = auto()
     DBRX         = auto()
     OLMO         = auto()
+    OLMOE        = auto()
     OPENELM      = auto()
     ARCTIC       = auto()
     DEEPSEEK2    = auto()
@@ -230,6 +234,7 @@ class MODEL_ARCH(IntEnum):
     JAIS         = auto()
     NEMOTRON     = auto()
     EXAONE       = auto()
+    GRANITE      = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -375,6 +380,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.COMMAND_R:      "command-r",
     MODEL_ARCH.DBRX:           "dbrx",
     MODEL_ARCH.OLMO:           "olmo",
+    MODEL_ARCH.OLMOE:          "olmoe",
     MODEL_ARCH.OPENELM:        "openelm",
     MODEL_ARCH.ARCTIC:         "arctic",
     MODEL_ARCH.DEEPSEEK2:      "deepseek2",
@@ -385,6 +391,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.JAIS:           "jais",
     MODEL_ARCH.NEMOTRON:       "nemotron",
     MODEL_ARCH.EXAONE:         "exaone",
+    MODEL_ARCH.GRANITE:        "granite",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -1027,6 +1034,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.OLMOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+    ],
     MODEL_ARCH.OPENELM: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -1205,6 +1229,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.GRANITE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 

gguf-py/gguf/gguf_writer.py

@@ -679,6 +679,12 @@ class GGUFWriter:
     def add_time_decay_extra_dim(self, dim: int) -> None:
         self.add_uint32(Keys.LLM.TIME_DECAY_EXTRA_DIM.format(arch=self.arch), dim)
 
+    def add_residual_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.RESIDUAL_SCALE.format(arch=self.arch), value)
+
+    def add_embedding_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.EMBEDDING_SCALE.format(arch=self.arch), value)
+
     def add_wkv_head_size(self, size: int) -> None:
         self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size)
 
@@ -703,6 +709,9 @@ class GGUFWriter:
     def add_sliding_window(self, value: int) -> None:
         self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value)
 
+    def add_attention_scale(self, value: float) -> None:
+        self.add_float32(Keys.Attention.SCALE.format(arch=self.arch), value)
+
     def add_pooling_type(self, value: PoolingType) -> None:
         self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)
 

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
+            "model.embed_tokens",                        # llama-hf nemotron olmoe
             "tok_embeddings",                            # llama-pth
             "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
@@ -54,7 +54,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
@@ -66,7 +66,7 @@ class TensorNameMap:
         MODEL_TENSOR.OUTPUT_NORM: (
             "gpt_neox.final_layer_norm",               # gptneox
             "transformer.ln_f",                        # gpt2 gpt-j falcon jais exaone
-            "model.norm",                              # llama-hf baichuan internlm2
+            "model.norm",                              # llama-hf baichuan internlm2 olmoe
             "norm",                                    # llama-pth
             "transformer.norm_f",                      # mpt dbrx
             "ln_f",                                    # refact bloom qwen gpt2
@@ -98,7 +98,7 @@ 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
+            "model.layers.{bid}.input_layernorm",                   # llama-hf nemotron olmoe
             "layers.{bid}.attention_norm",                          # llama-pth
             "language_model.encoder.layers.{bid}.input_layernorm",  # persimmon
             "model.layers.{bid}.ln1",                               # yi
@@ -142,7 +142,7 @@ class TensorNameMap:
 
         # Attention query
         MODEL_TENSOR.ATTN_Q: (
-            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron
+            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe
             "layers.{bid}.attention.wq",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.query",                  # bert
             "transformer.h.{bid}.attn.q_proj",                           # gpt-j
@@ -154,7 +154,7 @@ class TensorNameMap:
 
         # Attention key
         MODEL_TENSOR.ATTN_K: (
-            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron
+            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe
             "layers.{bid}.attention.wk",                               # llama-pth
             "encoder.layer.{bid}.attention.self.key",                  # bert
             "transformer.h.{bid}.attn.k_proj",                         # gpt-j
@@ -167,7 +167,7 @@ class TensorNameMap:
 
         # Attention value
         MODEL_TENSOR.ATTN_V: (
-            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron
+            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe
             "layers.{bid}.attention.wv",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.value",                  # bert
             "transformer.h.{bid}.attn.v_proj",                           # gpt-j
@@ -185,7 +185,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.attn.out_proj",                       # mpt
             "transformer.h.{bid}.self_attention.dense",                     # falcon
             "h.{bid}.self_attention.dense",                                 # bloom
-            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron
+            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe
             "layers.{bid}.attention.wo",                                    # llama-pth
             "encoder.layer.{bid}.attention.output.dense",                   # bert
             "transformer.h.{bid}.attn.out_proj",                            # gpt-j
@@ -229,7 +229,7 @@ class TensorNameMap:
             "transformer.h.{bid}.ln_2",                                      # gpt2 refact qwen jais exaone
             "h.{bid}.post_attention_layernorm",                              # bloom
             "transformer.blocks.{bid}.norm_2",                               # mpt
-            "model.layers.{bid}.post_attention_layernorm",                   # llama-hf nemotron
+            "model.layers.{bid}.post_attention_layernorm",                   # llama-hf nemotron olmoe
             "layers.{bid}.ffn_norm",                                         # llama-pth
             "language_model.encoder.layers.{bid}.post_attention_layernorm",  # persimmon
             "model.layers.{bid}.ln2",                                        # yi
@@ -253,7 +253,7 @@ class TensorNameMap:
         MODEL_TENSOR.FFN_GATE_INP: (
             "layers.{bid}.feed_forward.gate",             # mixtral
             "model.layers.{bid}.block_sparse_moe.gate",   # mixtral
-            "model.layers.{bid}.mlp.gate",                # qwen2moe
+            "model.layers.{bid}.mlp.gate",                # qwen2moe olmoe
             "transformer.decoder_layer.{bid}.router",     # Grok
             "transformer.blocks.{bid}.ffn.router.layer",  # dbrx
         ),
@@ -295,7 +295,7 @@ class TensorNameMap:
             "layers.{bid}.feed_forward.experts.w3",          # mixtral (merged)
             "transformer.decoder_layer.{bid}.moe.linear_v",  # Grok (merged)
             "transformer.blocks.{bid}.ffn.experts.mlp.v1",   # dbrx
-            "model.layers.{bid}.mlp.experts.up_proj",        # qwen2moe (merged)
+            "model.layers.{bid}.mlp.experts.up_proj",        # qwen2moe olmoe (merged)
         ),
 
         MODEL_TENSOR.FFN_UP_SHEXP: (
@@ -327,7 +327,7 @@ class TensorNameMap:
             "layers.{bid}.feed_forward.experts.w1",         # mixtral (merged)
             "transformer.decoder_layer.{bid}.moe.linear",   # Grok (merged)
             "transformer.blocks.{bid}.ffn.experts.mlp.w1",  # dbrx
-            "model.layers.{bid}.mlp.experts.gate_proj",     # qwen2moe (merged)
+            "model.layers.{bid}.mlp.experts.gate_proj",     # qwen2moe olmoe (merged)
         ),
 
         MODEL_TENSOR.FFN_GATE_SHEXP: (
@@ -367,7 +367,7 @@ class TensorNameMap:
             "layers.{bid}.feed_forward.experts.w2",          # mixtral (merged)
             "transformer.decoder_layer.{bid}.moe.linear_1",  # Grok (merged)
             "transformer.blocks.{bid}.ffn.experts.mlp.w2",   # dbrx
-            "model.layers.{bid}.mlp.experts.down_proj",      # qwen2moe (merged)
+            "model.layers.{bid}.mlp.experts.down_proj",      # qwen2moe olmoe (merged)
         ),
 
         MODEL_TENSOR.FFN_DOWN_SHEXP: (
@@ -378,7 +378,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.q_norm",                            # cohere
+            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
             "transformer.layers.{bid}.attn.q_norm",                           # openelm
@@ -387,7 +387,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.k_norm",                            # cohere
+            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
             "transformer.layers.{bid}.attn.k_norm",                           # openelm

grammars/README.md

@@ -120,7 +120,7 @@ You can use GBNF grammars:
 
 - In [llama-server](../examples/server):
     - For any completion endpoints, passed as the `json_schema` body field
-    - For the `/chat/completions` endpoint, passed inside the `response_format` body field (e.g. `{"type", "json_object", "schema": {"items": {}}}`)
+    - For the `/chat/completions` endpoint, passed inside the `response_format` body field (e.g. `{"type", "json_object", "schema": {"items": {}}}` or `{ type: "json_schema", json_schema: {"schema": ...} }`)
 - In [llama-cli](../examples/main), passed as the `--json` / `-j` flag
 - To convert to a grammar ahead of time:
     - in CLI, with [examples/json_schema_to_grammar.py](../examples/json_schema_to_grammar.py)

include/llama.h

@@ -441,6 +441,7 @@ extern "C" {
     LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
     LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
     LLAMA_API int32_t llama_n_layer    (const struct llama_model * model);
+    LLAMA_API int32_t llama_n_head     (const struct llama_model * model);
 
     LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
 

scripts/compare-commits.sh

@@ -8,6 +8,9 @@ fi
 set -e
 set -x
 
+# verify at the start that the compare script has all the necessary dependencies installed
+./scripts/compare-llama-bench.py --check
+
 bench_args="${@:3}"
 
 rm -f llama-bench.sqlite > /dev/null

scripts/compare-llama-bench.py

@@ -92,6 +92,7 @@ help_s = (
     "If the columns are manually specified, then the results for each unique combination of the "
     "specified values are averaged WITHOUT weighing by the --repetitions parameter of llama-bench."
 )
+parser.add_argument("--check", action="store_true", help="check if all required Python libraries are installed")
 parser.add_argument("-s", "--show", help=help_s)
 parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
 
@@ -99,6 +100,10 @@ known_args, unknown_args = parser.parse_known_args()
 
 logging.basicConfig(level=logging.DEBUG if known_args.verbose else logging.INFO)
 
+if known_args.check:
+    # Check if all required Python libraries are installed. Would have failed earlier if not.
+    sys.exit(0)
+
 if unknown_args:
     logger.error(f"Received unknown args: {unknown_args}.\n")
     parser.print_help()

src/llama-sampling.cpp

@@ -236,9 +236,10 @@ llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_conte
     const int n_vocab = llama_n_vocab(llama_get_model(ctx));
 
     // TODO: do not allocate each time
-    std::vector<llama_token_data> cur(n_vocab);
+    std::vector<llama_token_data> cur;
+    cur.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
+        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
     }
 
     llama_token_data_array cur_p = {

src/llama-vocab.cpp

@@ -50,7 +50,7 @@ struct naive_trie {
             res.first->second.insert(key + 1, len - 1, value);
         }
     }
-    std::pair<const char *, size_t> get_longest_prefix(const char * key, size_t len, size_t offset = 0) {
+    std::pair<const char *, size_t> get_longest_prefix(const char * key, size_t len, size_t offset = 0) const {
         if (len == 0 || offset == len) {
             return std::make_pair(key, offset);
         }
@@ -79,6 +79,15 @@ struct naive_trie {
 // impl
 //
 
+struct llm_tokenizer {
+   llm_tokenizer() {}
+   virtual ~llm_tokenizer() = default;
+};
+
+llama_vocab::~llama_vocab() {
+    delete tokenizer;
+}
+
 int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string & token_right) const {
     GGML_ASSERT(token_left.find(' ')   == std::string::npos);
     GGML_ASSERT(token_left.find('\n')  == std::string::npos);
@@ -187,10 +196,16 @@ struct llm_bigram_spm {
     size_t size;
 };
 
-struct llm_tokenizer_spm {
-    llm_tokenizer_spm(const llama_vocab & vocab) : vocab(vocab) {}
+struct llm_tokenizer_spm : llm_tokenizer {
+    llm_tokenizer_spm(const llama_vocab & /*vocab*/) : llm_tokenizer() {}
+};
+
+struct llm_tokenizer_spm_session {
+    llm_tokenizer_spm_session(const llama_vocab & vocab) : vocab(vocab),
+        spm_tokenizer(static_cast<const llm_tokenizer_spm *>(vocab.tokenizer)) {}
 
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+
         // split string into utf8 chars
         int index = 0;
         size_t offs = 0;
@@ -271,7 +286,7 @@ private:
             return;
         }
 
-        resegment(symbols[p->second.first],  output);
+        resegment(symbols[p->second.first], output);
         resegment(symbols[p->second.second], output);
     }
 
@@ -279,7 +294,6 @@ private:
         if (left == -1 || right == -1) {
             return;
         }
-
         const std::string text = std::string(symbols[left].text, symbols[left].n + symbols[right].n);
         auto token = vocab.token_to_id.find(text);
 
@@ -306,10 +320,10 @@ private:
     }
 
     const llama_vocab & vocab;
+    const llm_tokenizer_spm * spm_tokenizer; // currently unused
 
     std::vector<llm_symbol> symbols;
     llm_bigram_spm::queue work_queue;
-
     std::map<std::string, std::pair<int, int>> rev_merge;
 };
 
@@ -352,8 +366,8 @@ struct llm_bigram_bpe {
     size_t size;
 };
 
-struct llm_tokenizer_bpe {
-    llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {
+struct llm_tokenizer_bpe : llm_tokenizer {
+    llm_tokenizer_bpe(const llama_vocab & vocab) : llm_tokenizer() {
         GGML_ASSERT(vocab.type == LLAMA_VOCAB_TYPE_BPE);
         switch (vocab.type_pre) {
             case LLAMA_VOCAB_PRE_TYPE_LLAMA3:
@@ -462,7 +476,14 @@ struct llm_tokenizer_bpe {
         }
     }
 
-    void append(const llama_vocab::id token_id, std::vector<llama_vocab::id> & output) const {
+    std::vector<std::string> regex_exprs;
+};
+
+struct llm_tokenizer_bpe_session {
+    llm_tokenizer_bpe_session(const llama_vocab & vocab) : vocab(vocab),
+        bpe_tokenizer(static_cast<const llm_tokenizer_bpe *>(vocab.tokenizer)) {}
+
+    static void append(const llama_vocab::id token_id, std::vector<llama_vocab::id> & output)  {
         output.push_back(token_id);
     }
 
@@ -501,12 +522,11 @@ struct llm_tokenizer_bpe {
 
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         int final_prev_index = -1;
-
-        const auto word_collection = unicode_regex_split(text, regex_exprs);
+        const auto word_collection = unicode_regex_split(text, bpe_tokenizer->regex_exprs);
 
         symbols_final.clear();
 
-        for (auto & word : word_collection) {
+        for (const auto & word : word_collection) {
             work_queue = llm_bigram_bpe::queue();
             symbols.clear();
 
@@ -609,7 +629,6 @@ private:
         if (left == -1 || right == -1) {
             return;
         }
-
         std::string left_token  = std::string(symbols[left].text,  symbols[left].n);
         std::string right_token = std::string(symbols[right].text, symbols[right].n);
 
@@ -633,12 +652,10 @@ private:
     }
 
     const llama_vocab & vocab;
-
-    std::vector<std::string> regex_exprs;
+    const llm_tokenizer_bpe * bpe_tokenizer;
 
     std::vector<llm_symbol> symbols;
     std::vector<llm_symbol> symbols_final;
-
     llm_bigram_bpe::queue work_queue;
 };
 
@@ -646,15 +663,18 @@ private:
 // WPM tokenizer
 //
 
-struct llm_tokenizer_wpm {
-    llm_tokenizer_wpm(const llama_vocab & vocab): vocab(vocab) {}
+struct llm_tokenizer_wpm : llm_tokenizer {
+    llm_tokenizer_wpm(const llama_vocab & /*vocab*/) : llm_tokenizer() {}
+};
 
-    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) const {
+struct llm_tokenizer_wpm_session {
+    llm_tokenizer_wpm_session(const llama_vocab & vocab) : vocab(vocab),
+        wpm_tokenizer(static_cast<const llm_tokenizer_wpm *>(vocab.tokenizer)) {}
+
+    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         const auto & token_map = vocab.token_to_id;
-
         // normalize and split by whitespace
         std::vector<std::string> words = preprocess(text);
-
         // bos token prepended already
 
         // find the longest tokens that form the words
@@ -699,7 +719,7 @@ struct llm_tokenizer_wpm {
     }
 
     // TODO: reduce string copies by using cpts_offs array
-    std::vector<std::string> preprocess(const std::string & text) const {
+    static std::vector<std::string> preprocess(const std::string & text)  {
         const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
         std::vector<std::string> words(1, "");
 
@@ -751,15 +771,17 @@ struct llm_tokenizer_wpm {
             //(cpt >= 0xFF00  && cpt <= 0xFFEF);
     }
 
+private:
     const llama_vocab & vocab;
+    const llm_tokenizer_wpm * wpm_tokenizer;
 };
 
 //
 // UGM tokenizer
 //
 
-struct llm_tokenizer_ugm {
-    llm_tokenizer_ugm(const llama_vocab & vocab) : vocab(vocab) {
+struct llm_tokenizer_ugm : llm_tokenizer {
+    llm_tokenizer_ugm(const llama_vocab & vocab) : llm_tokenizer() {
         if (vocab.precompiled_charsmap.size() > 0) {
             size_t charsmap_offset = 0;
 
@@ -805,6 +827,30 @@ struct llm_tokenizer_ugm {
         unknown_token_score = min_score - unknown_token_score_penalty;
     }
 
+    // escaped space symbol - U+2581 (Lower One Eighth Block)
+    const std::string escaped_space = "\xE2\x96\x81";
+
+    const char * prefix_replacements = NULL;
+    size_t prefix_replacements_size = 0;
+
+    const uint32_t * xcda_array = NULL;
+    size_t xcda_array_size = 0;
+
+    struct naive_trie user_defined_token_matcher;
+
+    float min_score = FLT_MAX;
+    float max_score = -FLT_MAX;
+
+    float unknown_token_score_penalty = 10.0;
+    float unknown_token_score;
+
+    struct naive_trie token_matcher;
+};
+
+struct llm_tokenizer_ugm_session {
+    llm_tokenizer_ugm_session(const llama_vocab & vocab) : vocab(vocab),
+        ugm_tokenizer(static_cast<const llm_tokenizer_ugm *>(vocab.tokenizer)) {}
+
     /* This implementation is based on SentencePiece optimized Viterbi algorithm for
      * unigram language models. The general idea is to:
      * - move along the input sequence in steps of one UTF code point,
@@ -843,7 +889,7 @@ struct llm_tokenizer_ugm {
             // traverse the token matcher trie to find a matching token
             bool single_codepoint_token_found = false;
             const struct best_tokenization & current_best = tokenization_results[input_offset];
-            const struct naive_trie * node = token_matcher.traverse(normalized[prefix_offset++]);
+            const struct naive_trie * node = ugm_tokenizer->token_matcher.traverse(normalized[prefix_offset++]);
 
             while (prefix_offset <= input_len && node != NULL) {
                 // check if we found valid token in prefix
@@ -873,7 +919,7 @@ struct llm_tokenizer_ugm {
             // if we didn't find a valid token corresponding to the whole UTF code point
             // then use unknown token as the tokenization of this UTF code point
             if (!single_codepoint_token_found) {
-                const double challenger_score = current_best.score_sum + unknown_token_score;
+                const double challenger_score = current_best.score_sum + ugm_tokenizer->unknown_token_score;
                 prefix_offset = input_offset + n_utf8_code_units;
                 struct best_tokenization & current_champ = tokenization_results[prefix_offset];
                 if (challenger_score > current_champ.score_sum) {
@@ -905,7 +951,6 @@ struct llm_tokenizer_ugm {
     }
 
 private:
-    const llama_vocab & vocab;
 
     // helper structure for returning normalization results
     struct normalization_result {
@@ -918,7 +963,7 @@ private:
         normalized->clear();
         normalized->reserve(input.size() * 3);
 
-        const std::string space = vocab.tokenizer_escape_whitespaces ? escaped_space : " ";
+        const std::string space = vocab.tokenizer_escape_whitespaces ? ugm_tokenizer->escaped_space : " ";
 
         bool shall_prepend_space = !vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix;
         bool shall_append_space = vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix;
@@ -1000,13 +1045,21 @@ private:
         size_t xcda_array_size;
     };
 
+    // this structure stores the best tokenization so far at input_offset
+    struct best_tokenization {
+        llama_token token_id;
+        size_t input_offset;
+        float score_sum;
+    };
+
     struct normalization_result normalize_prefix(const std::string & input, size_t input_offset) {
         if (input_offset == input.size()) {
             return { &input[input_offset], 0, 0 };
         }
 
         // if input prefix matches some user-defined token return this token as normalization result
-        auto user_defined_token_match = user_defined_token_matcher.get_longest_prefix(&input[input_offset], input.size() - input_offset);
+        auto user_defined_token_match =
+           ugm_tokenizer->user_defined_token_matcher.get_longest_prefix(&input[input_offset], input.size() - input_offset);
         if (user_defined_token_match.second > 0) {
             return { &input[input_offset], user_defined_token_match.second, user_defined_token_match.second };
         }
@@ -1014,8 +1067,8 @@ private:
         size_t longest_prefix_length = 0;
         size_t longest_prefix_offset = 0;
 
-        if (xcda_array_size > 0) {
-            struct xcda_array_view xcda_view(xcda_array, xcda_array_size);
+        if (ugm_tokenizer->xcda_array_size > 0) {
+            struct xcda_array_view xcda_view(ugm_tokenizer->xcda_array, ugm_tokenizer->xcda_array_size);
 
             // Find the longest normalized sequence matching the input prefix by walking
             // the XOR-compressed compact double array (XCDA) starting from the root node
@@ -1051,50 +1104,27 @@ private:
 
         if (longest_prefix_length > 0) {
             // we have a match, so return the replacement sequence
-            if (longest_prefix_offset >= prefix_replacements_size) {
+            if (longest_prefix_offset >= ugm_tokenizer->prefix_replacements_size) {
                 throw std::runtime_error("Index out of array bounds in precompiled charsmap!");
             }
-            const char * prefix_replacement = &prefix_replacements[longest_prefix_offset];
+            const char * prefix_replacement = &(ugm_tokenizer->prefix_replacements)[longest_prefix_offset];
             return { prefix_replacement, strlen(prefix_replacement), longest_prefix_length };
-        } else {
-            // check if the input prefix contains a valid sequence of UTF-8 code units
-            try {
-                // if yes, return this sequence unmodified
-                size_t prefix_offset = input_offset;
-                unicode_cpt_from_utf8(input, prefix_offset);
-                return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset };
-            } catch (std::invalid_argument & /*ex*/) {
-                // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER
-                return { "\xEF\xBF\xBD", 3, 1 };
-            }
+        }
+
+        // check if the input prefix contains a valid sequence of UTF-8 code units
+        try {
+            // if yes, return this sequence unmodified
+            size_t prefix_offset = input_offset;
+            unicode_cpt_from_utf8(input, prefix_offset);
+            return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset };
+        } catch (std::invalid_argument & /*ex*/) {
+            // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER
+            return { "\xEF\xBF\xBD", 3, 1 };
         }
     }
 
-    // escaped space symbol - U+2581 (Lower One Eighth Block)
-    const std::string escaped_space = "\xE2\x96\x81";
-
-    const char * prefix_replacements = NULL;
-    size_t prefix_replacements_size = 0;
-
-    const uint32_t * xcda_array = NULL;
-    size_t xcda_array_size = 0;
-
-    struct naive_trie user_defined_token_matcher;
-
-    // this structure stores the best tokenization so far at input_offset
-    struct best_tokenization {
-        llama_token token_id;
-        size_t input_offset;
-        float score_sum;
-    };
-
-    float min_score = FLT_MAX;
-    float max_score = -FLT_MAX;
-
-    float unknown_token_score_penalty = 10.0;
-    float unknown_token_score;
-
-    struct naive_trie token_matcher;
+    const llama_vocab & vocab;
+    const llm_tokenizer_ugm * ugm_tokenizer;
 };
 
 //
@@ -1155,8 +1185,8 @@ static std::vector<uint8_t> llama_unescape_rwkv_token(const std::string & escape
     return output;
 }
 
-struct llm_tokenizer_rwkv {
-    llm_tokenizer_rwkv(const llama_vocab & vocab): vocab(vocab) {
+struct llm_tokenizer_rwkv : llm_tokenizer {
+    llm_tokenizer_rwkv(const llama_vocab & vocab) : llm_tokenizer() {
         // RWKV supports arbitrary byte tokens, but the vocab struct only supports string tokens.
         // For now, we decode the vocab here into the lookup we'll use for tokenization.
 
@@ -1168,11 +1198,17 @@ struct llm_tokenizer_rwkv {
         }
     }
 
+    struct naive_trie token_matcher;
+};
+
+struct llm_tokenizer_rwkv_session {
+    llm_tokenizer_rwkv_session(const llama_vocab & vocab) : vocab(vocab),
+        rwkv_tokenizer(static_cast<const llm_tokenizer_rwkv &>(*vocab.tokenizer)) {}
+
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         uint32_t position = 0;
-
         while (position < text.size()) {
-            const struct naive_trie * node = token_matcher.traverse(text[position]);
+            const struct naive_trie * node = rwkv_tokenizer.token_matcher.traverse(text[position]);
             if (node == NULL) {
                 // no matching token found, add unknown token
                 output.push_back(vocab.special_unk_id);
@@ -1197,11 +1233,33 @@ struct llm_tokenizer_rwkv {
         }
     }
 
+private:
     const llama_vocab & vocab;
-
-    struct naive_trie token_matcher;
+    const llm_tokenizer_rwkv & rwkv_tokenizer;
 };
 
+void llama_vocab::init_tokenizer() {
+    switch (type) {
+        case LLAMA_VOCAB_TYPE_SPM:
+            tokenizer = new llm_tokenizer_spm(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_BPE:
+            tokenizer = new llm_tokenizer_bpe(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_WPM:
+            tokenizer = new llm_tokenizer_wpm(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_UGM:
+            tokenizer = new llm_tokenizer_ugm(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_RWKV:
+            tokenizer = new llm_tokenizer_rwkv(*this);
+            break;
+        default:
+            GGML_ABORT("unsupported vocab type");
+    }
+}
+
 //
 // (de-) tokenize
 //
@@ -1263,7 +1321,7 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
 
             // if a fragment is text ( not yet processed )
             if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
-                auto & raw_text = fragment.raw_text;
+                const auto & raw_text = fragment.raw_text;
 
                 auto raw_text_base_offset = fragment.offset;
                 auto raw_text_base_length = fragment.length;
@@ -1362,7 +1420,13 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
     }
 }
 
-std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool add_special, bool parse_special) {
+std::vector<llama_vocab::id> llama_tokenize_internal(
+        const llama_vocab & vocab,
+        std::string raw_text,
+        bool add_special,
+        bool parse_special) {
+    GGML_ASSERT(vocab.tokenizer && "Tokenizer not initialized. Call llama_vocab::init_tokenizer() first.");
+
     std::vector<llama_vocab::id> output;
     std::forward_list<fragment_buffer_variant> fragment_buffer;
 
@@ -1399,9 +1463,9 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        llm_tokenizer_spm tokenizer(vocab);
                         llama_escape_whitespace(raw_text);
-                        tokenizer.tokenize(raw_text, output);
+                        llm_tokenizer_spm_session session(vocab);
+                        session.tokenize(raw_text, output);
                         is_prev_special = false;
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
@@ -1423,10 +1487,11 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
             } break;
         case LLAMA_VOCAB_TYPE_BPE:
             {
-                llm_tokenizer_bpe tokenizer(vocab);
-
+                llm_tokenizer_bpe_session session(vocab);
+                // it calls some other methods that are not exist in llm_tokenizer,
+                // here just cast it to bpe tokenizer object
                 if (add_special) {
-                    tokenizer.append_bos(output);
+                    session.append_bos(output);
                 }
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
@@ -1435,15 +1500,15 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
-                        tokenizer.append(fragment.token, output);
+                        session.append(fragment.token, output);
                     }
                 }
 
                 if (add_special) {
-                    tokenizer.append_eos(output);
-                    tokenizer.check_double_bos_eos(output);
+                    session.append_eos(output);
+                    session.check_double_bos_eos(output);
                 }
             } break;
         case LLAMA_VOCAB_TYPE_WPM:
@@ -1453,7 +1518,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
                     output.push_back(vocab.special_cls_id);
                 }
 
-                llm_tokenizer_wpm tokenizer(vocab);
+                llm_tokenizer_wpm_session session(vocab);
 
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
@@ -1462,7 +1527,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
                     }
@@ -1475,12 +1540,11 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
             } break;
         case LLAMA_VOCAB_TYPE_UGM:
             {
-                llm_tokenizer_ugm tokenizer(vocab);
-
                 if (add_special && vocab.tokenizer_add_bos != 0) {
                     GGML_ASSERT(vocab.special_bos_id != -1);
                     output.push_back(vocab.special_bos_id);
                 }
+                llm_tokenizer_ugm_session session(vocab);
 
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
@@ -1488,7 +1552,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
                     }
@@ -1508,6 +1572,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
             } break;
         case LLAMA_VOCAB_TYPE_RWKV:
             {
+                llm_tokenizer_rwkv_session session(vocab);
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
                         auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
@@ -1516,8 +1581,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
 
-                        llm_tokenizer_rwkv tokenizer(vocab);
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
                     }
@@ -1634,13 +1698,13 @@ llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
 }
 
 int32_t llama_tokenize_impl(
-    const struct llama_vocab & vocab,
-                  const char * text,
-                     int32_t   text_len,
-                 llama_token * tokens,
-                     int32_t   n_tokens_max,
-                        bool   add_special,
-                        bool   parse_special) {
+        const struct llama_vocab & vocab,
+                      const char * text,
+                         int32_t   text_len,
+                     llama_token * tokens,
+                         int32_t   n_tokens_max,
+                            bool   add_special,
+                            bool   parse_special) {
     auto res = llama_tokenize_internal(vocab, std::string(text, text_len), add_special, parse_special);
     if (n_tokens_max < (int) res.size()) {
         // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
@@ -1765,6 +1829,8 @@ int32_t llama_detokenize_impl(
                          int32_t   text_len_max,
                             bool   remove_special,
                             bool   unparse_special) {
+    GGML_ASSERT(vocab.tokenizer && "Tokenizer not initialized. Call llama_vocab::init_tokenizer() first.");
+
     int32_t avail = text_len_max;
     int32_t total = 0;
 

src/llama-vocab.h

@@ -7,6 +7,8 @@
 #include <unordered_map>
 #include <map>
 
+struct llm_tokenizer;
+
 struct llama_vocab {
     using id    = llama_token;
     using token = std::string;
@@ -61,7 +63,14 @@ struct llama_vocab {
 
     std::vector<char> precompiled_charsmap;
 
+    llm_tokenizer * tokenizer = nullptr;
+
+    llama_vocab() = default;
+    ~llama_vocab();
+
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const;
+
+    void init_tokenizer();
 };
 
 //

src/llama.cpp

@@ -202,6 +202,7 @@ enum llm_arch {
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_DBRX,
     LLM_ARCH_OLMO,
+    LLM_ARCH_OLMOE,
     LLM_ARCH_OPENELM,
     LLM_ARCH_ARCTIC,
     LLM_ARCH_DEEPSEEK2,
@@ -213,6 +214,7 @@ enum llm_arch {
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_EXAONE,
     LLM_ARCH_RWKV6,
+    LLM_ARCH_GRANITE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -251,6 +253,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COMMAND_R,       "command-r"    },
     { LLM_ARCH_DBRX,            "dbrx"         },
     { LLM_ARCH_OLMO,            "olmo"         },
+    { LLM_ARCH_OLMOE,           "olmoe"        },
     { LLM_ARCH_OPENELM,         "openelm"      },
     { LLM_ARCH_ARCTIC,          "arctic"       },
     { LLM_ARCH_DEEPSEEK2,       "deepseek2"    },
@@ -262,6 +265,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_NEMOTRON,        "nemotron"     },
     { LLM_ARCH_EXAONE,          "exaone"       },
     { LLM_ARCH_RWKV6,           "rwkv6"        },
+    { LLM_ARCH_GRANITE,         "granite"      },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -301,6 +305,8 @@ enum llm_kv {
     LLM_KV_RESCALE_EVERY_N_LAYERS,
     LLM_KV_TIME_MIX_EXTRA_DIM,
     LLM_KV_TIME_DECAY_EXTRA_DIM,
+    LLM_KV_RESIDUAL_SCALE,
+    LLM_KV_EMBEDDING_SCALE,
 
     LLM_KV_ATTENTION_HEAD_COUNT,
     LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -315,6 +321,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_KV_LORA_RANK,
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
+    LLM_KV_ATTENTION_SCALE,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_FREQ_BASE,
@@ -405,6 +412,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_RESCALE_EVERY_N_LAYERS,            "%s.rescale_every_n_layers"            },
     { LLM_KV_TIME_MIX_EXTRA_DIM,                "%s.time_mix_extra_dim"                },
     { LLM_KV_TIME_DECAY_EXTRA_DIM,              "%s.time_decay_extra_dim"              },
+    { LLM_KV_RESIDUAL_SCALE,                    "%s.residual_scale"                    },
+    { LLM_KV_EMBEDDING_SCALE,                   "%s.embedding_scale"                   },
 
     { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
     { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
@@ -419,6 +428,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"           },
     { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
     { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"         },
+    { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                  },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
@@ -1193,6 +1203,26 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_OLMOE,
+        {
+            { 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_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { 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" },
+        },
+    },
     {
         LLM_ARCH_OPENELM,
         {
@@ -1432,6 +1462,22 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_CHANNEL_MIX_RECEPTANCE,    "blk.%d.channel_mix_receptance" },
         },
     },
+    {
+        LLM_ARCH_GRANITE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { 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_UNKNOWN,
         {
@@ -2277,6 +2323,7 @@ enum e_model {
     MODEL_MEDIUM,
     MODEL_LARGE,
     MODEL_XL,
+    MODEL_A1_7B,
     MODEL_A2_7B,
     MODEL_8x7B,
     MODEL_8x22B,
@@ -2349,6 +2396,11 @@ struct llama_hparams {
     float f_max_alibi_bias = 0.0f;
     float f_logit_scale    = 0.0f;
 
+    // Additional scale factors (Granite)
+    float f_residual_scale  = 0.0f;
+    float f_embedding_scale = 0.0f;
+    float f_attention_scale = 0.0f;
+
     bool causal_attn   = true;
     bool use_alibi     = false;
     bool attn_soft_cap = false;
@@ -2411,6 +2463,9 @@ struct llama_hparams {
         if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true;
         if (!is_float_close(this->expert_weights_scale,  other.expert_weights_scale,  EPSILON)) return true;
         if (!is_float_close(this->rope_yarn_log_mul,     other.rope_yarn_log_mul,     EPSILON)) return true;
+        if (!is_float_close(this->f_residual_scale,      other.f_residual_scale,      EPSILON)) return true;
+        if (!is_float_close(this->f_embedding_scale,     other.f_embedding_scale,     EPSILON)) return true;
+        if (!is_float_close(this->f_attention_scale,     other.f_attention_scale,     EPSILON)) return true;
 
         return false;
     }
@@ -5241,6 +5296,7 @@ static const char * llama_model_type_name(e_model type) {
         case MODEL_MEDIUM:        return "0.4B";
         case MODEL_LARGE:         return "0.8B";
         case MODEL_XL:            return "1.5B";
+        case MODEL_A1_7B:         return "A1.7B";
         case MODEL_A2_7B:         return "A2.7B";
         case MODEL_8x7B:          return "8x7B";
         case MODEL_8x22B:         return "8x22B";
@@ -5791,6 +5847,14 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_OLMOE:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 16: model.type = e_model::MODEL_A1_7B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_OPENELM:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -5987,6 +6051,20 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GRANITE:
+            {
+                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);
+                ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale);
+                ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale);
+                ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale);
+
+                switch (hparams.n_layer) {
+                    case 40: model.type = e_model::MODEL_3B; break;
+                    // Add additional layer/vocab/etc checks here for other model sizes
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         default: (void)0;
     }
 
@@ -6029,8 +6107,15 @@ static void llm_load_vocab(
             vocab.special_mask_id = -1;
             vocab.linefeed_id     = -1;
 
+            // read vocab size from metadata
+            if (!ml.get_key(LLM_KV_VOCAB_SIZE, vocab.n_vocab, false)) {
+                vocab.n_vocab = 0;
+                LLAMA_LOG_WARN("%s: there is no vocab_size in metadata, vocab.n_vocab will be set to %u\n", __func__, vocab.n_vocab);
+            }
             return;
-        } else if (tokenizer_model == "llama") {
+        }
+
+        if (tokenizer_model == "llama") {
             vocab.type = LLAMA_VOCAB_TYPE_SPM;
 
             // default special tokens
@@ -6319,6 +6404,8 @@ static void llm_load_vocab(
     }
     GGML_ASSERT(vocab.id_to_token.size() == vocab.token_to_id.size());
 
+    vocab.init_tokenizer();
+
     // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n'
     if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
         // For Fill-In-the-Middle (FIM)/infill models which where converted
@@ -6368,11 +6455,11 @@ static void llm_load_vocab(
     } else if (vocab.type == LLAMA_VOCAB_TYPE_WPM) {
         vocab.linefeed_id = vocab.special_pad_id;
     } else if (vocab.type == LLAMA_VOCAB_TYPE_RWKV) {
-        const std::vector<int> ids = llama_tokenize_internal(vocab, "\n", false);
+        const std::vector<int> ids = llama_tokenize_internal(model.vocab, "\n", false);
         GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         vocab.linefeed_id = ids[0];
     } else {
-        const std::vector<int> ids = llama_tokenize_internal(vocab, "\xC4\x8A", false); // U+010A
+        const std::vector<int> ids = llama_tokenize_internal(model.vocab, "\xC4\x8A", false); // U+010A
         GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         vocab.linefeed_id = ids[0];
     }
@@ -6685,6 +6772,12 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
         LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
         LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
     }
+
+    if (model.arch == LLM_ARCH_GRANITE) {
+        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);
+    }
 }
 
 // Returns false if cancelled by progress_callback
@@ -6853,6 +6946,7 @@ static bool llm_load_tensors(
             case LLM_ARCH_LLAMA:
             case LLM_ARCH_REFACT:
             case LLM_ARCH_MINICPM:
+            case LLM_ARCH_GRANITE:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
 
@@ -8018,6 +8112,44 @@ static bool llm_load_tensors(
                         layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
                     }
                 } break;
+            case LLM_ARCH_OLMOE:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd});
+                        layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
+
+                        GGML_ASSERT(n_expert      > 0);
+                        GGML_ASSERT(n_expert_used > 0);
+
+                        // MoE branch
+                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert});
+                        layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert});
+                        layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert});
+                    }
+                } break;
             case LLM_ARCH_OPENELM:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
@@ -8798,6 +8930,11 @@ static struct ggml_tensor * llm_build_inp_embd(
         ggml_set_input(lctx.inp_embd);
     }
 
+    // For Granite architecture
+    if (hparams.f_embedding_scale != 0.0f) {
+        inpL = ggml_scale(ctx, inpL, hparams.f_embedding_scale);
+    }
+
     cb(inpL, "inp_embd", -1);
 
     return inpL;
@@ -9501,7 +9638,7 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(
         struct ggml_tensor * cur,
         struct ggml_tensor * x_prev,
         struct ggml_tensor ** wkv_state) {
-    size_t n_embed      = cur->ne[0];
+    size_t n_embd       = cur->ne[0];
     size_t n_seq_tokens = cur->ne[1];
     size_t n_seqs       = cur->ne[2];
 
@@ -9512,8 +9649,8 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(
 
     struct ggml_tensor * sx = ggml_sub(ctx, x_prev, cur);
 
-    sx  = ggml_reshape_2d(ctx, sx,  n_embed, n_tokens);
-    cur = ggml_reshape_2d(ctx, cur, n_embed, n_tokens);
+    sx  = ggml_reshape_2d(ctx, sx,  n_embd, n_tokens);
+    cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens);
 
     struct ggml_tensor * xxx = ggml_add(ctx, ggml_mul(ctx, sx, layer->time_mix_lerp_x), cur);
 
@@ -9538,11 +9675,11 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(
         xxx
     );
 
-    struct ggml_tensor *mw = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], 0);
-    struct ggml_tensor *mk = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * sizeof(float));
-    struct ggml_tensor *mv = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * 2 * sizeof(float));
-    struct ggml_tensor *mr = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * 3 * sizeof(float));
-    struct ggml_tensor *mg = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * 4 * sizeof(float));
+    struct ggml_tensor *mw = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], 0);
+    struct ggml_tensor *mk = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));
+    struct ggml_tensor *mv = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));
+    struct ggml_tensor *mr = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));
+    struct ggml_tensor *mg = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));
 
     struct ggml_tensor * xw = ggml_add(
         ctx,
@@ -9611,7 +9748,7 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(
         )
     );
 
-    w = ggml_add(ctx, w, ggml_reshape_1d(ctx, layer->time_mix_decay, n_embed));
+    w = ggml_add(ctx, w, ggml_reshape_1d(ctx, layer->time_mix_decay, n_embd));
     w = ggml_exp(ctx, ggml_neg(ctx, ggml_exp(ctx, w)));
     w = ggml_reshape_4d(ctx, w, 1, head_size, head_count, n_tokens);
 
@@ -9620,21 +9757,21 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(
     r = ggml_transpose(ctx, r);
 
     struct ggml_tensor * wkv_output = ggml_rwkv_wkv(ctx, k, v, r, layer->time_mix_first, w, *wkv_state);
-    cur = ggml_view_1d(ctx, wkv_output, n_embed * n_tokens, 0);
-    *wkv_state = ggml_view_1d(ctx, wkv_output, n_embed * head_size * n_seqs, n_embed * n_tokens * sizeof(float));
+    cur = ggml_view_1d(ctx, wkv_output, n_embd * n_tokens, 0);
+    *wkv_state = ggml_view_1d(ctx, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float));
 
     // group norm with head_count groups
-    cur = ggml_reshape_3d(ctx, cur, n_embed / head_count, head_count, n_tokens);
+    cur = ggml_reshape_3d(ctx, cur, n_embd / head_count, head_count, n_tokens);
     cur = ggml_norm(ctx, cur, 64e-5f);
 
     // Convert back to regular vectors.
-    cur = ggml_reshape_2d(ctx, cur, n_embed, n_tokens);
+    cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens);
     cur = ggml_add(ctx, ggml_mul(ctx, cur, layer->time_mix_ln), layer->time_mix_ln_b);
 
     cur = ggml_mul(ctx, cur, g);
     cur = llm_build_lora_mm(lctx, ctx, layer->time_mix_output, cur);
 
-    return ggml_reshape_3d(ctx, cur, n_embed, n_seq_tokens, n_seqs);
+    return ggml_reshape_3d(ctx, cur, n_embd, n_seq_tokens, n_seqs);
 }
 
 static struct ggml_tensor * llm_build_rwkv6_channel_mix(
@@ -10076,6 +10213,7 @@ struct llm_build_context {
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
 
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
         for (int il = 0; il < n_layer; ++il) {
             struct ggml_tensor * inpSA = inpL;
 
@@ -10128,7 +10266,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, lctx, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -10139,6 +10277,11 @@ struct llm_build_context {
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
 
+            // For Granite architecture
+            if (hparams.f_residual_scale) {
+                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+            }
+
             struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
             cb(ffn_inp, "ffn_inp", il);
 
@@ -10175,6 +10318,11 @@ struct llm_build_context {
                 cb(cur, "ffn_moe_out", il);
             }
 
+            // For Granite architecture
+            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);
 
@@ -10194,6 +10342,12 @@ struct llm_build_context {
 
         // lm_head
         cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        // For Granite architecture
+        if (hparams.f_logit_scale) {
+            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+        }
+
         cb(cur, "result_output", -1);
 
         ggml_build_forward_expand(gf, cur);
@@ -13832,6 +13986,134 @@ struct llm_build_context {
         return gf;
     }
 
+    // based on the build_qwen2moe() function, changes:
+    //   * removed shared experts
+    //   * removed bias
+    //   * added q, k norm
+    struct ggml_cgraph * build_olmoe() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(Kcur, "Kcur_normed", 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);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur_rope", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur_rope", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // MoE branch
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_moe_ffn(ctx0, lctx, 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,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, false,
+                    false, 0.0,
+                    cb, il);
+            cb(cur, "ffn_moe_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     struct ggml_cgraph * build_openelm() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
@@ -15591,6 +15873,7 @@ static struct ggml_cgraph * llama_build_graph(
 
     switch (model.arch) {
         case LLM_ARCH_LLAMA:
+        case LLM_ARCH_GRANITE:
             {
                 result = llm.build_llama();
             } break;
@@ -15712,6 +15995,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_olmo();
             } break;
+        case LLM_ARCH_OLMOE:
+            {
+                result = llm.build_olmoe();
+            } break;
         case LLM_ARCH_OPENELM:
             {
                 result = llm.build_openelm();
@@ -16375,7 +16662,7 @@ static int llama_decode_internal(
     const uint32_t n_tokens_all = batch_all.n_tokens;
 
     if (n_tokens_all == 0) {
-        LLAMA_LOG_ERROR("%s: n_tokens == 0", __func__);
+        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
@@ -16388,7 +16675,7 @@ static int llama_decode_internal(
     if (batch_all.token) {
         for (uint32_t i = 0; i < n_tokens_all; ++i) {
             if (batch_all.token[i] < 0 || (uint32_t)batch_all.token[i] >= model.vocab.n_vocab) {
-                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch_all.token[i]);
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch_all.token[i]);
                 return -1;
             }
         }
@@ -16676,7 +16963,7 @@ static int llama_encode_internal(
     const uint32_t n_tokens = batch.n_tokens;
 
     if (n_tokens == 0) {
-        LLAMA_LOG_ERROR("%s: n_tokens == 0", __func__);
+        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
@@ -16689,7 +16976,7 @@ static int llama_encode_internal(
     if (batch.token) {
         for (uint32_t i = 0; i < n_tokens; ++i) {
             if (batch.token[i] < 0 || (uint32_t)batch.token[i] >= model.vocab.n_vocab) {
-                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch.token[i]);
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
                 return -1;
             }
         }
@@ -18845,6 +19132,10 @@ int32_t llama_n_layer(const struct llama_model * model) {
     return model->hparams.n_layer;
 }
 
+int32_t llama_n_head(const struct llama_model * model) {
+    return model->hparams.n_head();
+}
+
 const struct llama_model * llama_get_model(const struct llama_context * ctx) {
     return &ctx->model;
 }
@@ -18883,6 +19174,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_ARCTIC:
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_CHATGLM:
+        case LLM_ARCH_GRANITE:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
@@ -18896,6 +19188,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_QWEN:
         case LLM_ARCH_QWEN2:
         case LLM_ARCH_QWEN2MOE:
+        case LLM_ARCH_OLMOE:
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3:
         case LLM_ARCH_GEMMA:

src/unicode.cpp

@@ -5,6 +5,7 @@
 #include "unicode.h"
 #include "unicode-data.h"
 
+#include <algorithm>
 #include <cassert>
 #include <cstddef>
 #include <cstdint>

tests/CMakeLists.txt

@@ -84,6 +84,25 @@ llama_test(test-tokenizer-0 NAME test-tokenizer-0-qwen2             ARGS ${CMAKE
 llama_test(test-tokenizer-0 NAME test-tokenizer-0-refact            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
 llama_test(test-tokenizer-0 NAME test-tokenizer-0-starcoder         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
 
+# build test-tokenizer-parallel target once and add many tests
+add_executable(test-tokenizer-parallel test-tokenizer-parallel.cpp)
+target_link_libraries(test-tokenizer-parallel PRIVATE common)
+install(TARGETS test-tokenizer-parallel RUNTIME)
+
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-bert-bge          ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bert-bge.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-command-r         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-command-r.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-deepseek-coder    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-coder.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-deepseek-llm      ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-llm.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-falcon            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-gpt-2             ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-2.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-llama-bpe         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-llama-spm         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-spm.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-mpt               ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-phi-3             ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-phi-3.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-qwen2             ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-qwen2.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-refact            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
+llama_test(test-tokenizer-parallel NAME test-tokenizer-parallel-starcoder         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
+
 # build test-tokenizer-1-bpe target once and add many tests
 add_executable(test-tokenizer-1-bpe test-tokenizer-1-bpe.cpp)
 target_link_libraries(test-tokenizer-1-bpe PRIVATE common)
@@ -119,6 +138,7 @@ llama_target_and_test(test-grammar-parser.cpp)
 llama_target_and_test(test-llama-grammar.cpp)
 llama_target_and_test(test-grammar-integration.cpp)
 llama_target_and_test(test-grad0.cpp)
+llama_target_and_test(test-barrier.cpp)
 # llama_target_and_test(test-opt.cpp) # SLOW
 llama_target_and_test(test-backend-ops.cpp)
 

tests/test-barrier.cpp

@@ -0,0 +1,93 @@
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <chrono>
+#include <iostream>
+#include <cstdio>
+#include <cstdlib>
+#include <cassert>
+#include <vector>
+
+#define MAX_NARGS 2
+
+int main(int argc, char *argv[]) {
+
+    int n_threads = 4;
+    int n_rounds  = 100;
+
+    if (argc > 1) {
+        n_threads = std::atoi(argv[1]);
+    }
+
+    if (argc > 2) {
+        n_rounds  = std::atoi(argv[2]);
+    }
+
+    struct ggml_init_params params = {
+        /* .mem_size   = */ 1024*1024*1024,
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ false,
+    };
+
+    struct ggml_context * ctx = ggml_init(params);
+
+    // Create graph
+    struct ggml_cgraph * gf = ggml_new_graph(ctx);
+
+    // Lots of small, parallel ops where barriers in between will dominate
+    struct ggml_tensor * out = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,  64);
+    for (int i = 0; i < 1000; i++) {
+        struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, 64, 128);
+        out = ggml_mul_mat(ctx, a, out);
+
+        struct ggml_tensor * d = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, 128, 64);
+        out = ggml_mul_mat(ctx, d, out);
+    }
+
+    ggml_build_forward_expand(gf, out);
+    int n_nodes = ggml_graph_n_nodes(gf);
+
+    // Create threadpool
+    struct ggml_threadpool_params tpp  = ggml_threadpool_params_default(n_threads);
+    struct ggml_threadpool* threadpool = ggml_threadpool_new(&tpp);
+    if (!threadpool) {
+        fprintf(stderr, "threadpool create failed : n_threads %d\n", n_threads);
+        exit(1);
+    }
+
+    // Create compute plan
+    struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads, threadpool);
+
+    std::vector<uint8_t> work_data(cplan.work_size);
+    cplan.work_data = work_data.data();
+
+    std::cerr << "graph-compute with"
+              << "\n n_threads: " << n_threads
+              << "\n   n_nodes: " << n_nodes
+              << "\n  n_rounds: " << n_rounds
+              << "\n";
+    // ggml_graph_print(gf);
+
+    // Warmup
+    ggml_graph_compute(gf, &cplan);
+
+    auto t0 = std::chrono::high_resolution_clock::now();
+
+    for (int i=0; i < n_rounds; i++) {
+        ggml_graph_compute(gf, &cplan);
+    }
+
+    auto t1 = std::chrono::high_resolution_clock::now();
+
+    auto usec = std::chrono::duration_cast<std::chrono::microseconds>(t1-t0).count();
+    auto nsec = std::chrono::duration_cast<std::chrono::nanoseconds>(t1-t0).count();
+    std::cerr << "graph-compute took " << usec << " usec "
+              << "\n " << (float) usec / n_rounds << " usec per-iter"
+              << "\n " << (float) nsec / (n_rounds * n_nodes) << " nsec per-node"
+              << "\n";
+
+    ggml_threadpool_free(threadpool);
+    ggml_free(ctx);
+
+    return 0;
+}

tests/test-tokenizer-parallel.cpp

@@ -0,0 +1,180 @@
+#include "llama.h"
+#include "common.h"
+#include "console.h"
+
+#include <cstdio>
+#include <string>
+#include <map>
+#include <vector>
+#include <fstream>
+#include <thread>
+
+using llama_tests = std::map<std::string, std::vector<llama_token>>;
+
+static llama_tests read_tests(const std::string & fname_inp, const std::string & fname_out) {
+    llama_tests tests;
+
+    std::ifstream ifs_inp(fname_inp);
+    if (!ifs_inp) {
+        fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_inp.c_str());
+        return tests;
+    }
+
+    std::string sraw((std::istreambuf_iterator<char>(ifs_inp)), std::istreambuf_iterator<char>());
+
+    std::ifstream ifs_out(fname_out);
+    if (!ifs_out) {
+        fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_out.c_str());
+        return tests;
+    }
+
+    std::vector<std::string> sout;
+    for (std::string line; std::getline(ifs_out, line);) {
+        sout.push_back(line);
+    }
+
+    const std::string sep = "\n__ggml_vocab_test__\n";
+
+    std::vector<std::string> sinp;
+
+    size_t pos = 0;
+    while (pos < sraw.size()) {
+        const size_t next = sraw.find(sep, pos);
+        if (next == std::string::npos) {
+            sinp.push_back(sraw.substr(pos));
+            break;
+        }
+        sinp.push_back(sraw.substr(pos, next - pos));
+        pos = next + sep.size();
+    }
+
+    if (sinp.size() != sout.size()) {
+        fprintf(stderr, "%s : error: input and output files have different number of tests\n", __func__);
+        return tests;
+    }
+
+    for (size_t i = 0; i < sinp.size(); ++i) {
+        const std::string & s = sinp[i];
+        const std::string & o = string_strip(sout[i]);
+
+        std::vector<llama_token> toks;
+
+        size_t pos = 0;
+        while (pos < o.size()) {
+            size_t next = o.find(' ', pos);
+            if (next == std::string::npos) {
+                next = o.size();
+            }
+            const std::string stok = o.substr(pos, next - pos);
+            toks.push_back(std::stoi(stok));
+            pos = next + 1;
+        }
+
+        tests[s] = toks;
+    }
+
+    return tests;
+}
+
+int main(int argc, char const *argv[]) {
+
+    if (argc < 2) {
+        fprintf(stderr, "Usage: %s vocab-file \n", argv[0]);
+        return 1;
+    }
+
+    const std::string fname = argv[1];
+
+    const std::string fname_inp = fname + ".inp";
+    const std::string fname_out = fname + ".out";
+
+    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
+
+    llama_model * model;
+    llama_context * ctx;
+
+    llama_backend_init();
+
+    // load the vocab
+    {
+        auto mparams = llama_model_default_params();
+
+        mparams.vocab_only = true;
+
+        model = llama_load_model_from_file(fname.c_str(), mparams);
+
+        if (model == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            return 1;
+        }
+
+        auto cparams = llama_context_default_params();
+
+        ctx = llama_new_context_with_model(model, cparams);
+
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            llama_free_model(model);
+            return 1;
+        }
+    }
+
+#ifdef _WIN32
+    // We need this for unicode console support
+    console::init(false, false);
+    atexit([]() { console::cleanup(); });
+#endif
+
+    const int nthread = std::thread::hardware_concurrency();
+    std::vector<std::thread> threads(nthread);
+
+    bool success = true;
+
+    const auto k_tests = [&]() -> llama_tests {
+        const auto res = read_tests(fname_inp, fname_out);
+
+        if (res.empty()) {
+            fprintf(stderr, "%s : error: no tests found\n", __func__);
+            exit(1);
+        }
+
+        return res;
+    }();
+
+    const bool add_special = false;
+
+    for (int i = 0; i < nthread; i++) {
+        threads[i] = std::thread([&]() {
+            for (const auto & test_kv : k_tests) {
+                const std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, add_special, false);
+
+                bool correct = res.size() == test_kv.second.size();
+                for (int i = 0; i < (int) res.size() && correct; ++i) {
+                    if (test_kv.second[i] != res[i]) {
+                        correct = false;
+                    }
+                }
+
+                if (!correct) {
+                    success = false;
+                }
+            }
+        });
+    }
+
+    for (int i = 0; i < nthread; i++) {
+        threads[i].join();
+    }
+
+    llama_free_model(model);
+    llama_free(ctx);
+
+    llama_backend_free();
+
+    printf("\n");
+    printf("Tests %s\n", success ? "passed" : "failed");
+
+    return success ? 0 : 3;
+}
+
+